BMC Neurosci 2016, 17(Suppl 1):54
DOI 10.1186/s12868-016-0283-6
BMC Neuroscience
Open Access
MEETING ABSTRACTS
25th Annual Computational Neuroscience
Meeting: CNS‑2016
Seogwipo City, Jeju-do, South Korea. 2–7 July 2016
Published: 18 August 2016
A1
Functional advantages of cell‑type heterogeneity in neural
circuits
Tatyana O. Sharpee1
1
Computational Neurobiology Laboratory, The Salk Institute for Biological
Studies, San Diego, CA, USA
Correspondence: Tatyana O. Sharpee ‑ sharpee@snl.salk.edu
BMC Neuroscience 2016, 17(Suppl 1):A1
Neural circuits are notorious for the complexity of their organization.
Part of this complexity is related to the number of different cell types
that work together to encode stimuli. I will discuss theoretical results
that point to functional advantages of splitting neural populations
into subtypes, both in feedforward and recurrent networks. These
results outline a framework for categorizing neuronal types based on
their functional properties. Such classification scheme could augment
classification schemes based on molecular, anatomical, and electrophysiological properties.
A2
Mesoscopic modeling of propagating waves in visual cortex
Alain Destexhe1,2
1
UNIC, CNRS, Gif sur Yvette, France; 2The European Institute for Theoretical
Neuroscience (EITN), Paris, France
Correspondence: Alain Destexhe ‑ destexhe@unic.cnrs‑gif.fr
BMC Neuroscience 2016, 17(Suppl 1):A2
Propagating waves are large-scale phenomena widely seen in the
nervous system, in both anesthetized and awake or sleeping states.
Recently, the presence of propagating waves at the scale of microns–
millimeters was demonstrated in the primary visual cortex (V1) of
macaque monkey. Using a combination of voltage-sensitive dye (VSD)
imaging in awake monkey V1 and model-based analysis, we showed
that virtually every visual input is followed by a propagating wave
(Muller et al., Nat Comm 2014). The wave was confined within V1, and
was consistent and repeatable for a given input. Interestingly, two
propagating waves always interact in a suppressive fashion, and sum
sublinearly. This is in agreement with the general suppressive effect
seen in other circumstances in V1 (Bair et al., J Neurosci 2003; Reynaud
et al., J Neurosci 2012).
To investigate possible mechanisms for this suppression we have
designed mean-field models to directly integrate the VSD experiments. Because the VSD signal is primarily caused by the summed
voltage of all membranes, it represents an ideal case for mean-field
models. However, usual mean-field models are based on neuronal
transfer functions such as the well-known sigmoid function, or functions estimated from very simple models. Any error in the transfer
function may result in wrong predictions by the corresponding meanfield model. To palliate this caveat, we have obtained semi-analytic
forms of the transfer function of more realistic neuron models. We
found that the same mathematical template can capture the transfer function for models such as the integrate-and-fire (IF) model, the
adaptive exponential (AdEx) model, up to Hodgkin–Huxley (HH) type
models, all with conductance-based inputs.
Using these transfer functions we have built “realistic” mean-field models for networks with two populations of neurons, the regular-spiking
(RS) excitatory neurons, showing spike frequency adaptation, and the
fast-spiking (FS) inhibitory neurons. This mean-field model can reproduce the propagating waves in V1, due to horizontal interactions, as
shown previously using IF networks. This mean-field model also reproduced the suppressive interactions between propagating waves. The
mechanism of suppression was based on the preferential recruitment
of inhibitory cells over excitatory cells by afferent activity, which acted
through the conductance-based shunting effect of the two waves
onto one another. The suppression was negligible in networks with
identical models for excitatory and inhibitory cells (such as IF networks). This suggests that the suppressive effect is a general phenomenon due to the higher excitability of inhibitory neurons in cortex, in
line with previous models (Ozeki et al., Neuron 2009).
Work done in collaboration with Yann Zerlaut (UNIC) for modeling,
Sandrine Chemla and Frederic Chavane (CNRS, Marseille) for in vivo
experiments. Supported by CNRS and the European Commission
(Human Brain Project).
A3
Dynamics and biomarkers of mental disorders
Mitsuo Kawato1
1
ATR Computational Neuroscience Laboratories, 2‑2 Hikaridai, Seika‑cho,
Soraku‑gun, Kyoto 619‑0288, Japan
Correspondence: Mitsuo Kawato ‑ kawato@hip.atr.co.jp
BMC Neuroscience 2016, 17(Suppl 1):A3
Current diagnoses of mental disorders are made in a categorical way,
as exemplified by DSM-5, but many difficulties have been encountered
in such categorical regimes: the high percentage of comorbidities,
usage of the same drug for multiple disorders, the lack of any validated animal model, and the situation where no epoch-making drug
has been developed in the past 30 years. NIMH started RDoC (research
domain criterion) to overcome these problems [1], and some successful results have been obtained, including common genetic risk loci [2]
and common neuroanatomical changes for multiple disorders [3] as
well as psychosis biotypes [4].
In contrast to the currently dominant molecular biology approach,
which basically assumes one-to-one mapping between genes and
disorders, I postulate the following dynamics-based view of psychiatric
disorders. Our brain is a nonlinear dynamical system that can generate
spontaneous spatiotemporal activities. The dynamical system is characterized by multiple stable attractors, only one of which corresponds
to a healthy or typically developed state. The others are pathological
states.
The most promising research approach within the above dynamical
view is to combine resting-state functional magnetic resonance imaging, machine learning, big data, and sophisticated neurofeedback.
© 2016 The Author(s). This article is distributed under the terms of the Creative Commons Attribution 4.0 International License
(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium,
provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license,
and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/
publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
BMC Neurosci 2016, 17(Suppl 1):54
Yahata et al. developed an ASD biomarker using only 16/9730 functional connections, and it did not generalize to MDD or ADHD but
moderately to schizophrenia [5]. Yamashita’s regression model of
working memory ability from functional connections [6] generalized
to schizophrenia and reproduced the severity of working-memory
deficits of four psychiatric disorders (in preparation).
With the further development of machine learning algorithms and
accumulation of reliable datasets, we hope to obtain a comprehensive
landscape of many psychiatric and neurodevelopmental disorders.
Guided by this full-spectrum structure, a tailor-made neurofeedback
therapy should be optimized for each patient [7].
References
1. Insel T, Cuthbert B, Garvey M., et al. Research domain criteria (RDoC):
toward a new classification framework for research on mental disorders.
Am J Psychiatry. 2010;167:748–51.
2. Cross-disorder group of the psychiatric genomics consortium: identification of risk loci with shared effects on five major psychiatric disorders: a
genome-wide analysis. Lancet. 2013;381:1371–9.
3. Goodkind M, et al. Identification of a common neurobiological substrate
for mental illness. JAMA Psychiatry. 2015;72:305–15.
4. Clementz BA, et al. Identification of distinct psychosis biotypes using
brain-based biomarkers. Am J Psychiatry. 2016;173:373–84.
5. Yahata N, Morimoto J, Hashimoto R, Lisi G, Shibata K, Kawakubo Y, Kuwabara H, Kuroda M, Yamada T, Megumi F, Imamizu H, Nanez JE, Takahashi
H, Okamoto Y, Kasai K, Kato N, Sasaki Y, Watanabe T, Kawato M: A small
number of abnormal brain connections predicts adult autism spectrum
disorder. Nature Commun. 2016;7:11254. doi:10.1038/ncomms11254.
6. Yamashita M, Kawato M, Imamizu H. Predicting learning plateau of working memory from whole-brain intrinsic network connectivity patterns. Sci
Rep. 2015;5(7622). doi:10.1038/srep07622.
7. ATR Brain Information Communication Research Laboratory Group.
DecNef
Project. Available at http://www.cns.atr.jp/decnefpro/ (2016).
F1
Precise recruitment of spiking output at theta frequencies
requires dendritic h‑channels in multi‑compartment models
of oriens‑lacunosum/moleculare hippocampal interneurons
Vladislav Sekulić1,2, Frances K. Skinner1,2,3
1
Krembil Research Institute, University Health Network, Toronto, Ontario,
Canada, M5T 2S8; 2Department of Physiology, University of Toronto,
Toronto, Ontario, Canada, M5S 1A8; 3 Department of Medicine
(Neurology), University of Toronto, Toronto, Ontario, Canada, M5T 2S8
Correspondence: Vladislav Sekulić ‑ vlad.sekulic@utoronto.ca
BMC Neuroscience 2016, 17(Suppl 1):F1
The theta rhythm (4–12 Hz) is a prominent network oscillation
observed in the mammalian hippocampus and is correlated with spatial navigation and mnemonic processing. Inhibitory interneurons of
the hippocampus fire action potentials at specific phases of the theta
rhythm, pointing to distinct functional roles of interneurons in shaping this rhythmic activity. One hippocampal interneuron type, the
oriens-lacunosum/moleculare (O-LM) cell, provides direct feedback
inhibition and regulation of pyramidal cell activity in the CA1 region.
O-LM cells express the hyperpolarization-activated, mixed-cation current (Ih) and, in vitro, demonstrate spontaneous firing at theta that is
impaired upon blockade of Ih. Work using dynamic clamp has shown
that in the presence of frequency-modulated artificial synaptic inputs,
O-LM cells exhibit a spiking resonance at theta frequencies that is not
dependent on Ih [1]. However, due to the somatic injection limitation
of dynamic clamp, the study could not examine the potential contributions of putative dendritic Ih or the integration of dendriticallylocated synaptic inputs. To overcome this, we have used a database of
previously developed multi-compartment computational models of
O-LM cells [2].
We situated our OLM cell models in an in vivo-like context by injecting Poisson-based synaptic background activities throughout their
dendritic arbors. Excitatory and inhibitory synaptic weights were
tuned to produce similar baseline activity prior to modulation of the
Page 2 of 112
inhibitory synaptic process at various frequencies (2–30 Hz). We found
that models with dendritic inputs expressed enhanced resonant firing at theta frequencies compared to models with somatic inputs. We
then performed detailed analyses on the outputs of the models with
dendritic inputs to further elucidate these results with respect to Ih distributions. The ability of the models to be recruited at the modulated
input frequencies was quantified using the rotation number, or average number of spikes across all input cycles. Models with somatodendritic Ih were recruited at >50 % of the input cycles for a wider range of
theta frequencies (3–9 Hz) compared to models with somatic Ih only
(3–4 Hz). Models with somatodendritic Ih also exhibited a wider range
of theta frequencies for which phase-locked output (vector strength
>0.75) was observed (4–12 Hz), compared to models with somatic Ih
(3–5 Hz). Finally, the phase of firing of models with somatodendritic Ih
given 8–10 Hz modulated input was delayed 180–230° relative to the
time of release from inhibitory synaptic input.
O-LM cells receive phasic inhibitory inputs at theta frequencies from
a subpopulation of parvalbumin-positive GABAergic interneurons in
the medial septum (MS) timed to the peak of hippocampal theta, as
measured in the stratum pyramidale layer [3]. Furthermore, O-LM cells
fire at the trough of hippocampal pyramidal layer theta in vivo [4], an
approximate 180˚ phase delay from the MS inputs, corresponding to
the phase delay in our models with somatodendritic Ih. Our results
suggest that, given dendritic synaptic inputs, O-LM cells require somatodendritic Ih channel expression to be precisely recruited during
the trough of hippocampal theta activity. Our strategy of leveraging
model databases that encompass experimental cell type-specificity
and variability allowed us to reveal critical biophysical factors that contribute to neuronal function within in vivo-like contexts.
Acknowledgements: Supported by NSERC of Canada, an Ontario
Graduate Scholarship, and the SciNet HPC Consortium.
References
1. Kispersky TJ, Fernandez FR, Economo MN, White JA. Spike resonance
properties in hippocampal O-LM cells are dependent on refractory
dynamics. J Neurosci. 2012;32(11):3637–51.
2. Sekulić V, Lawrence JJ, Skinner FK. Using multi-compartment ensemble
modeling as an investigative tool of spatially distributed biophysical balances: application to hippocampal oriens-lacunosum/moleculare (O-LM)
cells. PLOS One. 2014;9(10):e106567.
3. Borhegyi Z, Varga V, Szilágyi, Fabo D, Freund TF. Phase segregation of
medial septal GABAergic neurons during hippocampal theta activity. J
Neurosci. 2004;24(39):8470–9.
4. Varga C, Golshani P, Soltesz I. Frequency-invariant temporal ordering of
interneuronal discharges during hippocampal oscillations in awake mice.
Proc Natl Acad Sci USA. 2012;109(40):E2726–34.
F2
Kernel methods in reconstruction of current sources
from extracellular potentials for single cells and the whole brains
Daniel K. Wójcik1, Chaitanya Chintaluri1, Dorottya Cserpán2, Zoltán
Somogyvári2
1
Department of Neurophysiology, Nencki Institute of Experimental
Biology, Warsaw, Poland; 2Department of Theory, Wigner Research Centre
for Physics of the Hungarian Academy of Sciences, Budapest, H‑1121,
Hungary
Correspondence: Daniel K. Wójcik ‑ d.wojcik@nencki.gov.pl
BMC Neuroscience 2016, 17(Suppl 1):F2
Extracellular recordings of electric potential, with a century old history, remain a popular tool for investigations of brain activity on all
scales, from single neurons, through populations, to the whole brains,
in animals and humans, in vitro and in vivo [1]. The specific information available in the recording depends on the physical settings of the
system (brain + electrode). Smaller electrodes are usually more selective and are used to capture local information (spikes from single cells
or LFP from populations) while larger electrodes are used for subdural
recordings (on the cortex, ECoG), on the scalp (EEG) but also as depth
electrodes in humans (called SEEG). The advantages of extracellular
electric potential are the ease of recording and its stability. Its problem
BMC Neurosci 2016, 17(Suppl 1):54
is interpretation: since electric field is long range one can observe
neural activity several millimeters from its source [2–4]. As a consequence every recording reflects activity of many cells, populations and
regions, depending on which level we focus. One way to overcome
this problem is to reconstruct the distribution of current sources (CSD)
underlying the measurement [5], typically done to identify activity on
systems level from multiple LFP on regular grids [6].
We recently proposed a kernel-based method of CSD estimation from
multiple LFP recordings from arbitrarily placed probes (i.e. not necessarily on a grid) which we called kernel Current Source Density method
(kCSD) [7]. In this overview we present the original proposition as well
as two recent developments, skCSD (single cell kCSD) and kESI (kernel
Electrophysiological Source Imaging). skCSD assumes that we know
which part of the recorded signal comes from a given cell and we
have access to the morphology of the cell. This could be achieved by
patching a cell, driving it externally while recording the potential on a
multielectrode array, injecting a dye, and reconstructing the morphology. In this case we know that the sources must be located on the cell
and this information can be successfully used in estimation. In kESI we
consider simultaneous recordings with subdural ECoG (strip and grid
electrodes) and with depth electrodes (SEEG). Such recordings are
taken on some epileptic patients prepared for surgical removal of epileptogenic zone. When MR scan of the patient head is taken and the
positions of the electrodes are known as well as the brain’s shape, the
idea of kCSD can be used to bound the possible distribution of sources
facilitating localization of the foci.
Acknowledgements: Polish Ministry for Science and Higher Education (grant 2948/7.PR/2013/2), Hungarian Scientific Research
Fund (Grant OTKA K113147), National Science Centre, Poland (Grant
2015/17/B/ST7/04123).
References
1. Buzsáki G, Anastassiou CA, Koch C. The origin of extracellular fields and
currents—EEG, ECoG, LFP and spikes. Nat Rev Neurosci. 2012;13:407–20.
2. Hunt MJ, Falinska M, Łęski S, Wójcik DK, Kasicki S. Differential effects
produced by ketamine on oscillatory activity recorded in the rat hippocampus, dorsal striatum and nucleus accumbens. J Psychopharmacol.
2011;25:808–21.
3. Lindén H, Tetzlaff T, Potjans TC, Pettersen KH, Gruen S, Diesmann M, Einevoll GT. Modeling the spatial reach of the LFP. Neuron. 2011;72:859–72..
4. Łęski S, Lindén H, Tetzlaff T, Pettersen KH, Einevoll GT. Frequency dependence of signal power and spatial reach of the local field potential. PLoS
Comput Biol. 2013;9:e1003137.
5. Wójcik DK. Current source density (CSD) analysis. In: Jaeger D, Jung R,
editors. Encyclopedia of computational neuroscience. SpringerReference.
Berlin: Springer; 2013.
6. Mitzdorf U. Current source-density method and application in cat
cerebral cortex: investigation of evoked potentials and EEG phenomena.
Physiol Rev. 1985;65:37–100.
7. Potworowski J, Jakuczun W, Łęski S, Wójcik DK. Kernel current source
density method. Neural Comput. 2012;24:541–75.
F3
The synchronized periods depend on intracellular transcriptional
repression mechanisms in circadian clocks
Jae Kyoung Kim1, Zachary P. Kilpatrick2, Matthew R. Bennett3, Kresimir
Josić2,4
1
Department of Mathematical Sciences, KAIST, Daejoen 34141,
Republic of Korea; 2Department of Mathematics, University of Houston,
Houston, TX 77004, USA; 3Department of Biochemistry and Cell Biology
and Institute of Biosciences and Bioengineering, Rice University, Houston,
TX 77005, USA; 4Department of Biology and Biochemistry, University
of Houston, Houston, TX 77004, USA
Correspondence: Jae Kyoung Kim ‑ jaekkim@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):F2
In mammals, circadian (~24 h) rhythms are mainly regulated by a
master circadian clock located in the suprachiasmatic nucleus (SCN)
[1]. The SCN consists of ~20,000 neurons, each of which generates
own rhythms via intracellular transcriptional negative feedback loop
Page 3 of 112
involving PER-CRY and BMAL1-CLOCK. These individual rhythms of
each neuron are synchronized through intercellular coupling via
neurotransmitters including VIP [2]. In this talk, I will discuss that the
synchronized periods via coupling signal strongly depend on the
mechanism of intracellular transcription repression [3–4]. Specifically,
using mathematical modeling and phase response curve analysis, we
find that the synchronized period of SCN stays close to the population mean of cells’ intrinsic periods (~24 h) if transcriptional repression
occurs via protein sequestration. However, the synchronized period
is far from the population mean when repression occurs via Hill-type
regulation (e.g. phosphorylation-based repression). These results
reveal the novel relationship between two major functions of the
SCN-intracellular rhythm generation and intercellular synchronization
of rhythms. Furthermore, this relationship provides an explanation for
why the protein sequestration is commonly used in circadian clocks of
multicellular organisms, which have a coupled master clock, but not in
unicellular organisms [4].
Acknowledgements: This work was funded by the National Institutes
of Health, through the joint National Science Foundation/National
Institute of General Medical Sciences Mathematical Biology Program
grant No. R01GM104974 (to M.R.B. and K.J.), National Science Foundation grants Nos. DMS-1311755 (to Z.P.K.) and DMS-1122094 (to K.J.),
the Robert A. Welch Foundation grant No. C-1729 (to M.R.B.), National
Science Foundation grant No. DMS-0931642 to the Mathematical Biosciences Institute (to J.K.K.), KAIST Research Allowance Grant
G04150020 (to J.K.K) and the TJ Park Science Fellowship of POSCO TJ
Park Foundation G01160001 (to J.K.K).
References
1. Dibner C, Schibler U, Albrecht U. The mammalian circadian timing system:
organization and coordination of central and peripheral clocks. Annu Rev
Physiol. 2010;72:517–49.
2. Welsh DK, Takahashi JS, Kay SA. Suprachiasmatic nucleus: cell autonomy
and network properties. Annu Rev Physiol. 2010;72:551.
3. Kim JK, Kilpatrick ZP, Bennett MR, Josić K. Molecular mechanisms that
regulate the coupled period of the mammalian circadian clock. Biophys J.
2014;106(9):2071–81.
4. Kim JK. Protein sequestration vs Hill-type repression in circadian clock
models (in revision).
O1
Assessing irregularity and coordination of spiking‑bursting
rhythms in central pattern generators
Irene Elices1, David Arroyo1, Rafael Levi1,2, Francisco B. Rodriguez1, Pablo
Varona1
1
Grupo de Neurocomputación Biológica, Dpto. de Ingeniería Informática,
Escuela Politécnica Superior, Universidad Autónoma de Madrid, Spain;
2
Department of Biological Sciences, University of Southern California, CA,
USA
Correspondence: Irene Elices ‑ irene.elices@uam.es
BMC Neuroscience 2016, 17(Suppl 1):O1
Found in all nervous systems, central pattern generators (CPGs) are
neural circuits that produce flexible rhythmic motor patterns. Their
robust and highly coordinated spatio-temporal activity is generated in
the absence of rhythmic input. Several invertebrate CPGs are among
the best known neural circuits, as their neurons and connections have
been identified and mapped. The crustacean pyloric CPG is one of
these flagship neural networks [1, 2]. Experimental and computational
studies of CPGs typically examine their rhythmic output in periodic
spiking-bursting regimes. Aiming to understand the fast rhythm negotiation of CPG neurons, here we present experimental and theoretical
analyses of the pyloric CPG activity in situations where irregular yet
coordinated rhythms are produced. In particular, we focus our study
in the context of two sources of rhythm irregularity: intrinsic damage
in the preparation, and irregularity induced by ethanol. The analysis of
non-periodic regimes can unveil important properties of the robust
dynamics controlling rhythm coordination in this system.
Adult male and female shore crabs (Carcinus maenas) were used for the
experimental recordings. The isolated stomatrogastric ganglion was kept
BMC Neurosci 2016, 17(Suppl 1):54
in Carcinus maenas saline. Membrane potentials were recorded intracellularly from the LP and PD cells, two mutually inhibitory neurons that form
a half-center oscillator in the pyloric CPG. Extracellular electrodes allowed
monitoring the overall CPG rhythm. Conductance-based models of the
pyloric CPG neurons and their associated graded synapses as described in
[3, 4] were also used in this dual experimental and theoretical study.
Irregularity and coordination of the CPG rhythms were analyzed using
measures characterizing the cells’ instantaneous waveform, period, duty
cycle, plateau, hyperpolarization and temporal structure of the spiking
activity, as well as measures describing instantaneous phases among
neurons in the irregular rhythms and their variability. Our results illustrate the strong robustness of the circuit to keep LP/PD phase relationships in intrinsic and induced irregularity conditions while allowing
a large variety of burst waveforms, durations and hyperpolarization
periods in these neurons. In spite of being electrically coupled to the
pacemaker cell of the circuit, the PD neurons showed a wide flexibility
to participate with larger burst durations in the CPG rhythm (and larger
increase in variability), while the LP neuron was more restricted in sustaining long bursts in the conditions analyzed. The conductance-based
models were used to explain the role of asymmetry in the dynamics
of the neurons and synapses to shape the irregular activity observed
experimentally. Taking into account the overall experimental and model
analyses, we discuss the presence of preserved relationships in the nonperiodic but coordinated bursting activity of the pyloric CPG, and their
role in the fast rhythm negotiating properties of this circuit.
Acknowledgements: We acknowledge support from MINECO
DPI2015-65833-P, TIN2014-54580-R, TIN-2012-30883 and ONRG grant
N62909-14-1-N279.
References
1. Marder E, Calabrese RL. Principles of rhythmic motor pattern generation.
Physiol Rev. 1996;76:687–717.
2. Selverston AI, Rabinovich MI, Abarbanel HDI, Elson R, Szücs A, Pinto RD,
Huerta R, Varona P. Reliable circuits from irregular neurons: a dynamical approach to understanding central pattern generators. J Physiol.
2000;94:357–74.
3. Latorre R, Rodríguez FB, Varona P. Neural signatures: multiple coding in
spiking-bursting cells. Biol Cybern. 2006;95:169–83.
4. Elices I, Varona P. Closed-loop control of a minimal central pattern generator network. Neurocomputing. 2015;170:55–62.
O2
Regulation of top‑down processing by cortically‑projecting
parvalbumin positive neurons in basal forebrain
Eunjin Hwang1, Bowon Kim1,2, Hio‑Been Han1,3, Tae Kim4, James T.
McKenna5, Ritchie E. Brown5, Robert W. McCarley5, Jee Hyun Choi1,2
1
Center for Neuroscience, Korea Institute of Science and Technology,
Hwarang‑ro 14‑gil 5, Seongbuk‑gu, Seoul 02792, South Korea;
2
Department of Neuroscience, University of Science and Technology,
217 Gajeong‑ro, Yuseong‑gu, Daejon 34113, South Korea; 3Department
of Psychology, Yonsei University, 50 Yonsei‑ro, Seodaemun‑gu, Seoul
03722, South Korea; 4Department of Psychiatry, Kyung Hee University
Hospital at Gangdong, 892, Dongnam‑ro, Gangdong‑gu, Seoul 05278,
South Korea; 5Department of Psychiatry, Veterans Administration Boston
Healthcare System and Harvard Medical School, Brockton, MA 02301, USA
Correspondence: Jee Hyun Choi ‑ jeechoi@kist.re.kr
BMC Neuroscience 2016, 17(Suppl 1):O2
Particular behaviors are associated with different spatio-temporal patterns of cortical EEG oscillations. A recent study suggests that the cortically-projecting, parvalbumin-positive (PV+) inhibitory neurons in
the basal forebrain (BF) play an important role in the state-dependent
control of cortical oscillations, especially ~40 Hz gamma oscillations
[1]. However, the cortical topography of the gamma oscillations which
are controlled by BF PV+ neurons and their relationship to behavior
are unknown. Thus, in this study, we investigated the spatio-temporal
patterns and the functional role of the cortical oscillations induced or
entrained by BF PV+ neurons by combining optogenetic stimulation
Page 4 of 112
of BF PV+ neurons with high-density EEG [2, 3] in channelrhodopsin-2
(ChR2) transduced PV-cre mice. First, we recorded the spatio-temporal
responses in the cortex with respect to the stimulation of BF PV+ neurons at various frequencies. The topographic response patterns were
distinctively different depending on the stimulation frequencies, and
most importantly, stimulation of BF PV+ neurons at 40 Hz (gamma
band frequency) induced a preferential enhancement of gamma
band oscillations in prefrontal cortex (PFC) with a statistically significant increase in intracortical connectivity within PFC. Second, optogenetic stimulation of BF PV+ neurons was applied while the mice were
exposed to auditory stimuli (AS) at 40 Hz. The time delay between
optogenetic stimulation and AS was tested and the phase response to
the AS was characterized. We found that the phase responses to the
click sound in PFC were modulated by the optogenetic stimulation
of BF PV+ neurons. More specifically, the advanced activation of BF
PV+ neurons by π/2 (6.25 ms) with respect to AS sharpened the phase
response to AS in PFC, while the anti-phasic activation (π, 12.5 ms)
blunted the phase response. Interestingly, like PFC, the primary auditory cortex (A1) also showed sharpened phase response for the π/2
advanced optogenetic BF PV+ neuron activation during AS. Considering that no direct influence of BF PV+ neurons on A1 was apparent in
the response to stimulation of BF PV+ neurons alone, the sharpened
phase response curve of A1 suggests a top-down influence of the
PFC. This result implies that the BF PV+ neurons may participate in
regulating the top-down influence that PFC exerts on primary sensory
cortices during attentive behaviors, and supports the idea that the
modulating activities of BF PV+ neurons might be a potential target
for restoring top-down cognitive functions as well as abnormal frontal
gamma oscillations associated with psychiatric disorders.
Acknowledgements: This research was supported by the Department of Veterans Affairs, the Korean National Research Council of
Science & Technology (No. CRC-15-04-KIST), NIMH R01 MH039683
and Basic Science Research Program through the National Research
Foundation of Korea (NRF) funded by the Ministry of Education
(2015R1D1A1A01059119). The contents of this report do not represent the views of the US Department of Veterans Affairs or the United
States government.
References
1. Kim T, et al. Cortically projecting basal forebrain parvalbumin neurons regulate cortical gamma band oscillations. Proc Natl Acad Sci.
2015;112(11):3535–40.
2. Choi JH, et al. High resolution electroencephalography in freely moving
mice. J Neurophysiol .2010;104(3):1825–34.
3. Lee M, et al. High-density EEG recordings of the freely moving mice using
polyimide-based microelectrode. J Vis Exp. 2011;47. http://www.jove.
com/details.php?id=2562. doi:10.3791/2562.
O3
Modeling auditory stream segregation, build‑up and bistability
James Rankin1, Pamela Osborn Popp1, John Rinzel1,2
1
Center for Neural Science, New York University, New York 10003, NY;
2
Courant Institute of Mathematical Sciences, New York University, New
York 10012, NY
Correspondence: James Rankin ‑ james.rankin@nyu.edu
BMC Neuroscience 2016, 17(Suppl 1):O3
With neuromechanistic modelling and psychoacoustic experiments we
study the perceptual dynamics of auditory streaming (cocktail party
problem). The stimulus is a sequence of two interleaved tones, A and
B in a repeating triplet pattern: ABA_ABA_ (‘_’ is a silent gap). Initially,
subjects hear a single integrated pattern, but after some seconds they
hear segregated A_A_A_ and _B___B__ streams (build-up of streaming
segregation). For long presentations, build-up is followed by irregular
alternations between integrated and segregated (auditory bistability).
We recently presented [1] the first neuromechanistic model of auditory
bistability; it incorporates common competition mechanisms of mutual
inhibition, slow adaptation and noise [2]. Our competition network is
BMC Neurosci 2016, 17(Suppl 1):54
Page 5 of 112
Germany; 3Bernstein Center for Computational Neuroscience,
Heidelberg‑Mannheim, Baden‑Württemberg, Germany
Correspondence: Alejandro Tabas ‑ atabas@bournemouth.ac.uk
†
Equal contribution
BMC Neuroscience 2016, 17(Suppl 1):O4
Fig. 1 A Model schematic: tone inputs IA and IB elicit pulsatile
responses in A1, which are pooled as inputs to a three-population
competition network. Central unit AB encodes integrated, peripheral
units A and B encode segregated. Mutual inhibition between units
and recurrent excitation are incorporated with adaptation and noise.
B A1 inputs show early initial adaptation, also if a pause is present.
Build-up function shows proportion segregated increasing over time,
here shown for three tone-frequency differences, DF, with no pause
(dashed) or with a pause (solid curves). Time-snapshots from model
(filled circles) agree with data (empty circles with SEM error bars, N = 8)
formulated to reside downstream of primary auditory cortex (A1). Neural responses in macaque A1 to triplet sequences [3] encode stimulus
features and provide the inputs to our network (Fig. 1A). In our model
recurrent excitation with an NMDA-like timescale links responses across
gaps between tones and between triplets. It captures the dynamics of
perceptual alternations and the stimulus feature dependence of percept durations. To account for build-up we incorporate early adaptation
of A1 responses [3] (Fig. 1B, upper). Early responses in A1 are broadly
tuned and do not reflect the frequency difference between the tones;
later responses show a clear tonotopic dependence. This adaptation
biases the initial percept towards integration, but occurs faster (~0.5 s)
than the gradual build-up process (~5–10 s). The low initial probability of segregation gradually builds up to the stable probability of later
bistable alternations (Fig. 1B, lower). During build-up, a pause in presentation may cause partial reset to integrated [4]. Our extended model
shows this behavior assuming that after a pause A1 responses recover
on the timescale of early adaptation. Moreover, the modeling results
agree with our psychoacoustic experiments (compare filled and open
circles in Fig. 1B, lower).
Conclusions For the first time, we offer an explanation of the discrepancy in the timescales of early A1 responses and the more gradual buildup process. Recovery of A1 responses can explain resetting for stimulus
pauses. Our model offers, to date, the most complete account of the
early and late dynamics for auditory streaming in the triplet paradigm.
References
1. Rankin J, Sussman E, Rinzel J. Neuromechanistic model of auditory bistability. PLoS Comput Biol. 2015;11:e1004555.
2. Shpiro A, Moreno-Bote R, Rubin N, Rinzel J. Balance between noise and
adaptation in competition models of perceptual bistability. J Comp
Neurosci. 2009;27:37–54.
3. Micheyl C, Tian B, Carlyon R, Rauschecker J. Perceptual organization of
tone sequences in the auditory cortex of awake macaques. Neuron.
2005;48:139–48.
4. Beauvois MW, Meddis R. Time decay of auditory stream biasing. Percept
Psychophys. 1997;59:81–6.
O4
Strong competition between tonotopic neural ensembles
explains pitch‑related dynamics of auditory cortex evoked fields
Alejandro Tabas1, André Rupp2,†, Emili Balaguer‑Ballester1,3,†
1
Faculty of Science and Technology, Bournemouth University,
Bournemouth, England, UK; 2Heidelberg University, Baden‑Württemberg,
Auditory evoked fields (AEFs) observed in MEG experiments systematically present a transient deflection known as the N100 m, elicited
around 100 ms after the tone onset in the antero-lateral Heschl’s Gyrus.
The exact N100m’s latency is correlated with the perceived pitch of a
wide range of stimulus [1, 2], suggesting that the transient component
reflects the processing of pitch in auditory cortex. However, the biophysical substrate of such precise relationship remains an enigma. Existing
models of pitch, focused on perceptual phenomena, did not explain the
mechanism generating cortical evoked fields during pitch processing
in biophysical detail. In this work, we introduce a model of interacting
neural ensembles describing, for the first time to our knowledge, how
cortical pitch processing gives rise to observed human neuromagnetic
responses and why its latency strongly correlates with pitch.
To provide a realistic cortical input, we used a recent model of the auditory periphery and realistic subcortical processing stages. Subcortical
processing was based on a delay-and-multiply operation carried out in
cochlear nucleus and inferior colliculus [3], resulting in realistic patterns
of neural activation in response to the stimulus periodicities. Subcortical activation is transformed into a tonotopic receptive-field-like representation [4] by a novel cortical circuit composed by functional blocks
characterised by a best frequency. Each block consist of an excitatory
and an inhibitory population, modelled using mean-field approximations [5]. Blocks interact with each other through local AMPA- and
NMDA-driven excitation and GABA-driven global inhibition [5].
The excitation-inhibition competition of the cortical model describes a
general pitch processing mechanism that explains the N100m deflection as a transient state in the cortical dynamics. The deflection is rapidly triggered by a rise in the activity elicited by the subcortical input,
peaks after the inhibition overcomes the input, and stabilises when
model dynamics reach equilibrium, around 100 ms after onset. As a
direct consequence of the connectivity structure among blocks, the
time necessary for the system to reach equilibrium depends on the
encoded pitch of the tone. The model quantitatively predicts observed
latencies of the N100m in agreement with available empirical data
[1, 2] in a series of stimuli (see Fig. 2), suggesting that the mechanism
potentially accounts for the N100 m dynamics.
References
1. Seither-Preisler A, Patterson R, Krumbholz K, Seither S, Lütkenhöner B.
Evidence of pitch processing in the N100 m component of the auditory
evoked field. Hear Res. 2006;213(1–2):88–98.
2. Roberts TP, Ferrari P, Stufflebeam SM, Poeppel D. Latency of the auditory
evoked neuromagnetic field components: stimulus dependence and
insights toward perception. J Clin Neurophysiol. 2000;17(2):114–29.
3. Meddis R, O’Mard LP. Virtual pitch in a computational physiological
model. J Acoust Soc Am. 2006;6:3861–9.
4. Balaguer-Ballester E, Clark, N. Understanding pitch perception as a
hierarchical process with top-down modulation. PLoS Comput Biol.
2009;5(3):e1000301.
5. Wong K-F, Wang X-J. A recurrent network mechanism of time integration
in perceptual decisions. J Neurosci. 2006;26(4):1314–28.
Fig. 2 N100 m predictions in comparison with available data [1, 2] for
a range of pure tones (A) and HCTs (B)
BMC Neurosci 2016, 17(Suppl 1):54
O5
A simple model of retinal response to multi‑electrode stimulation
Matias I. Maturana1,2, David B. Grayden2,3, Shaun L. Cloherty4, Tatiana
Kameneva2, Michael R. Ibbotson1,5, Hamish Meffin1,5
1
National Vision Research Institute, Australian College of Optometry, 3053,
Australia; 2NeuroEngineering Laboratory, Dept. Electrical & Electronic
Eng., University of Melbourne, 3010, Australia; 3Centre for Neural
Engineering, University of Melbourne, 3010, Australia; 4Department
of Physiology, Monash University, 3800, Australia; 5ARC Centre
of Excellence for Integrative Brain Function, Department Optometry
and Vision Sciences, University of Melbourne, 3010, Australia
Correspondence: Hamish Meffin ‑ hmeffin@unimelb.edu.au
BMC Neuroscience 2016, 17(Suppl 1):O5
Retinal implants can restore vision to patients suffering photoreceptor
loss by stimulating surviving retinal ganglion cells (RGCs) via an array
of microelectrodes implanted within the eye [1]. However, the acuity offered by existing devices is low, limiting the benefits to patients.
Improvements may come by increasing the number of electrodes
in new devices and providing patterned vision, which necessitates
stimulation using multiple electrodes simultaneously. However, simultaneous stimulation poses a number of problems due to cross-talk
between electrodes and uncertainty regarding the resulting activation
pattern.
Here, we present a model and methods for estimating the responses of
RGCs to simultaneous electrical stimulation. Whole cell in vitro patch
clamp recordings were obtained from 25 RGCs with various morphological types in rat retina. The retinae were placed onto an array of
20 stimulating electrodes. Biphasic current pulses with 500 µs phase
duration and 50 µs interphase gap were applied simultaneously to all
electrodes at a frequency of 10 Hz, with the amplitude of current on
each electrode sampled independently from a Gaussian distribution.
A linear-nonlinear model was fit to the responses of each RGC using
spike-triggered covariance analyses on 80 % of the recorded data.
The analysis revealed a single significant principle component corresponding to the electrical receptive field for each cell, with the second
largest principle component having negligible effect on the neural
response (Fig. 3a). This indicates that interactions between electrodes
are approximately linear in their influence on the cells’ responses.
Furthermore, the spike-triggered ensemble showed two clusters (red
and blue in Fig. 3a) corresponding to stimulation that had a net effect
that was either anodic first or cathodic first. The electrical receptive
fields for both anodic first and cathodic first stimulation were highly
similar (Fig. 3b). They consisted of a small number (1–4) of electrodes
that were close to the cell body (green dot).
Fig. 3 a Spike triggered covariance showing the full set of stimuli
(black dots) projected onto the first two principle components.
Stimuli causing a spike formed two clusters: net cathodic first pulses
(blue) and net anodic first pulse (red). b Electrical receptive fields
superimposed on the electrode array are shown for the cathodic first
(blue) and anodic first clusters (red)
Page 6 of 112
The remaining 20 % of data were used to validate the model. The
average model prediction root-mean-square error was 7 % over the
25 cells. The accuracy of the model indicates that the linear-nonlinear
model is appropriate to describe the responses of RGCs to electrical
stimulation.
Acknowledgements: This research was supported by the Australian Research Council (ARC). MI, HM, and SC acknowledge support
through the Centre of Excellence for Integrative Brain Function
(CE140100007), TK through ARC Discovery Early Career Researcher
Award (DE120102210) and HM and TK through the ARC Discovery Projects funding scheme (DP140104533).
Reference
1. Hadjinicolaou AE, Meffin H, Maturana M, Cloherty SL, Ibbotson MR.
Prosthetic vision: devices, patient outcomes and retinal research. Clin Exp
Optom. 2015;98(5):395–410.
O6
Noise correlations in V4
area correlate with behavioral performance in visual
discrimination task
Veronika Koren1,2, Timm Lochmann1,2, Valentin Dragoi3, Klaus
Obermayer1,2
1
Institute of Software Engineering and Theoretical Computer Science,
Technische Universitaet Berlin, Berlin, 10587, Germany; 2 Bernstein Center
for Computational Neuroscience Berlin, Humboldt‑Universitaet zu Berlin,
Berlin, 10115, Germany; 3Department of Neurobiology and Anatomy,
University of Texas‑Houston Medical School, Houston, TX 77030, USA
Correspondence: Veronika Koren ‑ veronika.koren@bccn‑berlin.de
BMC Neuroscience 2016, 17(Suppl 1):O6
Linking sensory coding and behavior is a fundamental question in
neuroscience. We have addressed this issue in behaving monkey visual cortex (areas V1 and V4) while animals were trained to perform a
visual discrimination task in which two successive images were either
rotated with respect to each other or were the same. We hypothesized that the animal’s performance in the visual discrimination task
depends on the quality of stimulus coding in visual cortex. We tested
this hypothesis by investigating the functional relevance of neuronal
correlations in areas V1 and V4 in relation to behavioral performance.
We measured two types of correlations: noise (spike count) correlations and correlations in spike timing. Surprisingly, both methods
showed that correct responses are associated with significantly higher
correlations in V4, but not V1, during the delay period between the
two stimuli. This suggests that pair-wise interactions during the spontaneous activity preceding the arrival of the stimulus sets the stage for
subsequent stimulus processing and importantly influences behavioral performance.
Experiments were conducted in 2 adult monkeys that were previously trained for the task. After 300 ms of fixation, the target stimulus, consisting of a naturalistic stimulus, is shown for 300 ms, and
after a random delay period (500–1200 ms), a test stimulus is shown
for 300 ms. The test can either be identical to the target stimulus
(match) or rotated with respect to the target (non-match). Monkey
responded by pressing a button and was rewarded for a correct
response with fruit juice. Two linear arrays with 16 recording channels each were used to record population activity in areas V1 and V4.
The difficulty of the task is calibrated individually to have 70 % correct responses on average. The analysis is conducted on non-match
condition, comparing activity in trials with correct responses with trials where the monkey responded incorrectly. Noise correlations were
assessed as pair-wise correlations of spike counts (method 1) and
of spike timing (method 2). For method 1, z-scores of spike counts
of binned spike trains are computed in individual trials. r_sc is computed as Pearson correlation coefficient of z-scores in all available
trials, balanced across correct/incorrect condition. For the method
2, cross-correlograms were computed, from which the cross-correlograms from shuffled trials are subtracted. Resulting function was
summed around zero lag and normalized with sum of autocorrelograms [1].
BMC Neurosci 2016, 17(Suppl 1):54
While firing rates of single units or of the population did not significantly change for correct and incorrect responses, noise correlations
during the delay period were significantly higher in V4 pairs, computed with both r_sc method (p = 0.0005 in monkey 1, sign-rank
test) and with r_ccg method (p = 0.0001 and p = 0.0280 in monkey 1
and 2, respectively, 50 ms integration window). This result is robust to
changes in the length of the bin (method 1) and to the length of the
summation window (method 2). In agreement with [2], we confirm the
importance of spontaneous activity preceding the stimulus on performance and suggest that higher correlations in V4 might be beneficial
for successful read-out and reliable transmission of the information
downstream.
References
1. Bair W, Zohary E, Newsome WT. Correlated firing in macaque visual
area MT: time scales and relationship to behavior. J Neurosci. 2001;
21(5):1676–97.
2. Gutnisky DA, Beaman CB, Lew SE, Dragoi V. Spontaneous fluctuations in
visual cortical responses influence population coding accuracy. Cereb
Cortex. 2016;1–19.
3. Cohen MR, Maunsell JH. Attention improves performance primarily by reducing interneuronal correlations. Nat Neurosci.
2009;12(12):1594–1600.
4. Nienborg HR, Cohen MR, Cumming BG. Decision-related activity in sensory neurons: correlations among neurons and with behavior. Annu Rev
Neurosci. 2012;35:463–83.
O7
Input‑location dependent gain modulation in cerebellar nucleus
neurons
Maria Psarrou1, Maria Schilstra1, Neil Davey1, Benjamin Torben‑Nielsen1,
Volker Steuber1
Centre for Computer Science and Informatics Research, University
of Hertfordshire, Hatfield, AL10 9AB, UK
Correspondence: Maria Psarrou ‑ m.psarrou@herts.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):O7
Gain modulation is a brain-wide principle of neuronal computation
that describes how neurons integrate inputs from different presynaptic sources. A gain change is a multiplicative operation that is defined
as a change in the sensitivity (or slope of the response amplitude) of
a neuron to one set of inputs (driving input) which results from the
activity of a second set of inputs (modulatory input) [1, 2].
Different cellular and network mechanisms have been proposed
to underlie gain modulation [2–4]. It is well established that input
features such as synaptic noise and plasticity can contribute to multiplicative gain changes [2–4]. However, the effect of neuronal morphology on gain modulation is relatively unexplored. Neuronal inputs
to the soma and dendrites are integrated in a different manner: whilst
dendritic saturation can introduce a strong non-linear relationship
between dendritic excitation and somatic depolarization, the relationship between somatic excitation and depolarization is more linear. The
non-linear integration of dendritic inputs can enhance the multiplicative effect of shunting inhibition in the presence of noise [3].
Neurons in the cerebellar nuclei (CN) provide the main gateway
from the cerebellum to the rest of the brain. Understanding how
inhibitory inputs from cerebellar Purkinje cells interact with excitatory inputs from mossy fibres to control output from the CN is at
the center of understanding cerebellar computation. In the present
study, we investigated the effect of inhibitory modulatory input on
CN neuronal output when the excitatory driving input was delivered
at different locations in the CN neuron. We used a morphologically
realistic conductance based CN neuron model [5] and examined the
change in output gain in the presence of distributed inhibitory input
under two conditions: (a) when the excitatory input was confined to
one compartment (the soma or a dendritic compartment) and, (b),
when the excitatory input was distributed across particular dendritic
regions at different distances from the soma. For both of these conditions, our results show that the arithmetic operation performed by
inhibitory synaptic input depends on the location of the excitatory
Page 7 of 112
synaptic input. In the presence of distal dendritic excitatory inputs,
the inhibitory input has a multiplicative effect on the CN neuronal
output. In contrast, excitatory inputs at the soma or proximal dendrites close to the soma undergo additive operations in the presence
of inhibitory input. Moreover, the amount of the multiplicative gain
change correlates with the distance of the excitatory inputs from the
soma, with increasing distances from the soma resulting in increased
gain changes and decreased additive shifts along the input axis. These
results indicate that the location of synaptic inputs affects in a systematic way whether the input undergoes a multiplicative or additive
operation.
References
1. Salinas E, Sejnowski TJ. Gain modulation in the central nervous system:
where behavior, neurophysiology, and computation meet. Neuroscientist. 2001;7(5):430–40.
2. Silver RA. Neuronal arithmetic. Nat Rev Neurosci. 2010;11(7):474–89.
3. Prescott SA, De Koninck Y. Gain control of firing rate by shunting inhibition: roles of synaptic noise and dendritic saturation. Proc Natl Acad Sci
USA. 2003;100(4):2076–81.
4. Rothman J, Cathala L, Steuber V, Silver RA. Synaptic depression enables
neuronal gain control. Nature. 2009;475:1015–18.
5. Steuber V, Schultheiss NW, Silver RA, De Schutter E, Jaeger D. Determinants of synaptic integration and heterogeneity in rebound firing
explored with data-driven models of deep cerebellar nucleus cells. J
Comput Neurosci. 2011;30(3):633–58.
O8
Analytic solution of cable energy function for cortical axons
and dendrites
Huiwen Ju1, Jiao Yu2, Michael L. Hines3, Liang Chen4 and Yuguo Yu1
1
School of Life Science and the Collaborative Innovation Center for Brain
Science, Fudan University, Shanghai, 200438, China; 2Linyi Hospital
of Traditional Chinese Medicine, 211 Jiefang Road, Lanshan, Linyi,
Shandong Province, 276000, China; 3Department of Neuroscience, Yale
University School of Medicine, New Haven, CT 06520, USA; 4Department
of Neurosurgery, Huashan Hospital, Shanghai Medical College, Fudan
University, Shanghai, China
Correspondence: Yuguo Yu ‑ yuyuguo@fudan.edu.cn
BMC Neuroscience 2016, 17(Suppl 1):O8
Accurate estimation of action potential (AP)-related metabolic cost
is essential for understanding energetic constraints on brain connections and signaling processes. Most previous energy estimates of the
AP were obtained using the Na+-counting method [1, 2], which seriously limits accurate assessment of metabolic cost of ionic currents
that underlie AP generation. Moreover, the effects of axonal geometry
and ion channel distribution on energy consumption related to AP
propagation have not been systematically investigated.
To address these issues, we return to the cable theory [3] that underlies our HH-type cortical axon model [4], which was constructed based
on experimental measurements. Based on the cable equation that
describes how ion currents flow along the cable as well as analysis of
the electrochemical energy in the equivalent circuit, we derived the
electrochemical energy function for the cable model,
∂2E
1
∂V
= INa (V − VNa ) + IK (V − VK ) + IL (V − VL ) −
ia
∂x∂t
2π a ∂x
max m3 h(V (x, t) − V )2 + g max n4 (V (x, t) − V )2
= gNa
K
Na
K
2
∂V
+ gL (V (x, t) − VL )2 + Ga
∂x
max
2
(5–100 mS/cm2), and
where gmax
Na (in a range of 50–650 mS/cm ), gK
gL = 0.033 mS/cm2 are the maximal sodium, maximal potassium, and
leak conductance per unit membrane area, respectively; and VNa = 60,
VK = −90 VL = −70 mV are the reversal potentials of the sodium,
potassium, and leak channels, respectively. The gate variables m, h,
and n are dimensionless activation and inactivation variables, which
describe the activation and inactivation processes of the sodium and
BMC Neurosci 2016, 17(Suppl 1):54
potassium channels [4]. This equation describes the AP-related energy
consumption rate per unit membrane area (cm2/s) at any axonal distance and any time. The individual terms on the right-hand side of the
equation represent the contributions of the sodium, potassium, leak,
and axial currents, respectively. Then we employed the cable energy
function to calculate energy consumption for unbranched axons and
axons with several degrees of branching (branching level, BL). Calculations based on this function distinguish between the contributions of
each item toward total energy consumption.
Our analytical approach predicts an inhomogeneous distribution of
metabolic cost along an axon with either uniformly or nonuniformly
distributed ion channels. The results show that the Na+-counting
method severely underestimates energy cost in the cable model by
20–70 %. AP propagation along axons that differ in length may require
over 15 % more energy per unit of axon area than that required by a
point model. However, actual energy cost can vary greatly depending
on axonal branching complexity, ion channel density distributions,
and AP conduction states. We also infer that the metabolic rate (i.e.
energy consumption rate) of cortical axonal branches as a function of
spatial volume exhibits a 3/4 power law relationship.
Acknowledgements: Dr. Yu thanks for the support from the National
Natural Science Foundation of China (31271170, 31571070), Shanghai program of Professor of Special Appointment (Eastern Scholar
SHH1140004).
References
1. Alle H, Roth A, Geiger JR. Energy-efficient action potentials in hippocampal mossy fibers. Science. 2009;325(5946):1405–8.
2. Carter BC, Bean BP. Sodium entry during action potentials of mammalian
neurons: incomplete inactivation and reduced metabolic efficiency in
fast-spiking neurons. Neuron. 2009;64(6):898–909.
3. Rall W. Cable theory for dendritic neurons. In: Methods in neuronal modeling. MIT Press; 1989. p. 9–92.
4. Yu Y, Hill AP, McCormick DA. Warm body temperature facilitates energy
efficient cortical action potentials. PLoS Comput Biol. 2012;8(4):e1002456.
O9
C. elegans interactome: interactive visualization of Caenorhabditis
elegans worm neuronal network
Jimin Kim1, Will Leahy2, Eli Shlizerman1,3
1
Department of Applied Mathematics, University of Washington, Seattle,
WA 98195, USA; 2Amazon.com Inc., Seattle, WA 98108, USA; 3Department
of Electrical Engineering, University of Washington, Seattle, WA 98195,
USA
Correspondence: Eli Shlizerman ‑ shlizee@uw.edu
BMC Neuroscience 2016, 17(Suppl 1):O9
Modeling neuronal systems involves incorporating the two layers: a
static map of neural connections (connectome), and biophysical processes that describe neural responses and interactions. Such a model
is called the ‘dynome’ of a neuronal system as it integrates a dynamical
system with the static connectome. Being closer to reproducing the
activity of a neuronal system, investigation of the dynome has more
potential to reveal neuronal pathways of the network than the static
connectome [1]. However, since the two layers of the dynome are
considered simultaneously, novel tools have to be developed for the
dynome studies. Here we present a visualization methodology, called
`interactome’, that allows to explore the dynome of a neuronal system
interactively and in real-time, by viewing the dynamics overlaid on a
graph representation of the connectome.
We apply our methodology to the nervous system of Caenorhabditis
elegans (C. elegans) worm, which connectome is almost fully resolved
[2], and a computational model of neural dynamics and interactions (gap and synaptic) based on biophysical experimental findings
was recently introduced [3]. Integrated together, C. elegans dynome
Page 8 of 112
Fig. 4 A Visualization of C. elegans dynome, B communication diagram between the dynome and the layout, C snapshots of visualization of C. elegans during the PLM/AVB excitations (forward crawling)
defines a unique set of neural dynamics of the worm. To visualize the
dynome, we propose a dynamic force-directed graph layout of the
connectome. The layout is implemented using D3 visualization platform [4], and is designed to communicate with an integrator of the
dynome. The two-way communication protocol between the layout
and the integrator allows for stimulating (injecting current) into any
subset of neurons at any time point (Fig. 4B). It also allows for simultaneously viewing the response of the network on top of the layout visualized by resizing graph nodes (neurons) according to their voltage. In
addition, we support structural changes in the connectome, such as
ablation of neurons and connections.
Our visualization and communication protocols thereby display the
stimulated network in an interactive manner and permit to explore different regimes that the stimulations induce. Indeed, with the interactome we are able to recreate various experimental scenarios, such as
stimulation of forward crawling (PLM/AVB neurons and/or ablation of
AVB) and show that its visualization assists in identifying patterns of
neurons in the stimulated network. As connectomes and dynomes of
additional neuronal systems are being resolved, the interactome will
enable exploring their functionality and inference to its underlying
neural pathways [5].
References
1. Kopell NJ, Gritton HJ, Whittingon MA, Kramer MA. Beyond the connectome: the dynome. Neuron. 2014;83(6):1319–28.
2. Varshney LR, Chen BL, Paniagua E, Hall DH, Chkolvski DB. Structural properties of the caenorhabditis elegans neuronal network. PLoS Comput
Biol. 2011;7(2):e1001066.
3. Kunert J, Shlizerman E, Kutz JN. Low-dimensional functionality of complex network dynamics: neurosensory integration in the Caenorhabditis
elegans connectome. Phys Rev E. 2014;89(5):052805.
4. Bostock M, Ogievetsky V, Heer J. D3 data-driven documents. IEEE.
2011;17(12):2301–9.
5. Kim J, Leahy W, Shlizerman E. C. elegans interactome: interactive visualization of Caenorhabditis elegans worm neuronal network. 2016 (in
submission).
O10
Is the model any good? Objective criteria for computational
neuroscience model selection
Justas Birgiolas1, Richard C. Gerkin1, Sharon M. Crook1,2
1
School of Life Science, Arizona State University, Tempe, AZ 85287, USA;
2
School of Mathematical and Statistical Sciences, Arizona State University,
Tempe, AZ, 85287, USA
Correspondence: Justas Birgiolas ‑ justas@asu.edu
BMC Neuroscience 2016, 17(Suppl 1):O10
Objectively evaluating and selecting computational models of biological neurons is an ongoing challenge in the field. Models vary in
morphological detail, channel mechanisms, and synaptic transmission
implementations. We present the results of an automated method for
evaluating computational models against property values obtained
BMC Neurosci 2016, 17(Suppl 1):54
Page 9 of 112
4.
5.
Tripathy SJ, Savitskaya J, Burton SD, Urban NN, Gerkin RC. NeuroElectro: a
window to the world’s neuron electrophysiology data. Front Neuroinform. 2014;8.
Chen WR, Shen GY, Shepherd GM, Hines ML, Midtgaard J. Multiple modes
of action potential initiation and propagation in mitral cell primary dendrite. J Neurophysiol. 2002;88(5):2755–64.
O11
Cooperation and competition of gamma oscillation mechanisms
Atthaphon Viriyopase1,2,3, Raoul‑Martin Memmesheimer1,3,4, and Stan
Gielen1,2
1
Donders Institute for Brain, Cognition and Behaviour, Radboud
University Nijmegen (Medical Centre), The Netherlands; 2Department
for Biophysics, Faculty of Science, Radboud University Nijmegen, The
Netherlands; 3Department for Neuroinformatics, Faculty of Science,
Radboud University Nijmegen, The Netherlands; 4Center for Theoretical
Neuroscience, Columbia University, New York, NY, USA
Correspondence: Atthaphon Viriyopase ‑ a.viriyopase@science.ru.nl
BMC Neuroscience 2016, 17(Suppl 1):O11
Fig. 5 The average deviations of models and cell electrophysiology
properties as measured in multiples of the 95 % CI bounds of experimental data means. Dashed line represents 1 CI bound threshold. Top
rows show average deviations across all models for each cell property.
Bottom rows show deviations across all cell properties for each model
from published cell electrophysiology studies. Seven published deterministic models of olfactory bulb mitral cells were selected from ModelDB [1] and simulated using NEURON’s Python interface [2]. Passive
and spike properties in response to step current stimulation pulses
were computed using the NeuronUnit [3] package and compared to
their respective, experimentally obtained means of olfactory bulb
mitral cell properties found in the NeuroElectro database [4].
Results reveal that across all models, the resting potential and input
resistance property means deviated the most from their experimentally measured means (Rinput t test p = 0.02, Vrest Wilcoxon-test
p = 0.01). The time constant, spike half-width, spike amplitude, and
spike threshold properties, in the order of decreasing average deviation, matched well with experimental data (p > 0.05) (Fig. 5 top).
In three models, the property deviations were, on average, outside the
95 % CI of the experimental means (Fig. 5 bottom), but these averages
were not significant (t test p > 0.05). All other models were within the
95 % CI, while the model of Chen et al. had the lowest deviation [5].
Overall, the majority of these olfactory bulb mitral cell models display
some properties that are not significantly different from their experimental means. However, the resting potential and input resistance
properties significantly differ from the experimental values. We demonstrate that NeuronUnit provides an objective method for evaluating
the fitness of computational neuroscience cell models against publicly
available data.
Acknowledgements: The work of JB, RG, and SMC was supported in
part by R01MH1006674 from the National Institutes of Health.
References
1. Hines ML, Morse T, Migliore M, Carnevale NT, Shepherd GM. ModelDB: a
database to support computational neuroscience. J Comput Neurosci.
2004;17(1):7–11.
2. Hines M, Davison AP, Muller E. NEURON and Python. Front Neuroinform.
2009;3:1.
3. Omar C, Aldrich J, Gerkin RC. Collaborative infrastructure for test-driven
scientific model validation. In: Companion proceedings of the 36th
international conference on software engineering. ACM; 2014. p. 524–7.
Two major mechanisms that underlie gamma oscillations are
InterNeuronal Gamma (“ING”), which is related to tonic excitation of
reciprocally coupled inhibitory interneurons (I-cells), and Pyramidal
InternNeuron Gamma (“PING”), which is mediated by coupled populations of excitatory pyramidal cells (E-cells) and I-cells. ING and PING
are thought to serve different biological functions. Using computer
simulations and analytical methods, we [1] therefore investigate which
mechanism (ING or PING) will dominate the dynamics of a network
when ING and PING interact and how the dominant mechanism may
switch.
We find that ING and PING oscillations compete: The mechanism
generating the higher oscillation frequency “wins”. It determines
the frequency of the network oscillations and suppresses the other
mechanism. The network oscillation frequency (green lines corresponding to the network topology given in Fig. 6C) corresponding to
Fig. 6 Oscillations in full and reduced networks of reciprocally
coupled pyramidal cells and interneurons. A, B Illustrate topologies of
reduced networks that generate “pure” ING and “pure” PING, respectively, while C highlights the topology of a “full” network that could
in principle generate either ING or PING oscillations or mixtures of
both. D, E Frequency of pure ING-rhythm generated by the reduced
network in A (blue line), pure PING-rhythm generated by the reduced
network in b (red line), and rhythms generated by the full network in
C (green line) as a function of mean current to I-cells I0,I and as function of mean current to E-cells I0,E, respectively. D Results for networks
with type-I interneurons while E shows results for networks with
type-II interneurons. Pyramidal cells are modeled as type-I Hodgkin–
Huxley neurons
BMC Neurosci 2016, 17(Suppl 1):54
the network with type-I-phase-response-curve interneurons and typeII-phase-response-curve interneurons is plotted in Fig. 6D, E, respectively. We explain our simulation results by a theoretical model that
allows a full theoretical analysis.
Our study suggests experimental approaches to decide whether
oscillatory activity in networks of interacting excitatory and inhibitory neurons is dominated by ING or PING oscillations and whether
the participating interneurons belong to class I or II. Consider as
an example networks with type-I interneurons where the external
drive to the E-cells, I0,E, is kept constant while the external drive to
the I-cells, I0,I, is varied. For both ING and PING dominated oscillations the frequency of the rhythm increases when I0,I increases (cf.
Fig. 6D). Observing such an increase does therefore not allow to
determine the underlying mechanism. However, the absolute value
of the first derivative of the frequency with respect to I0,I allows a
distinction, as it is much smaller for PING than for ING (cf. Fig. 6D).
In networks with type-II interneurons, the non-monotonic dependence near the ING-PING transition may be a characteristic hallmark
to detect the oscillation character (and the interneuron type):
Decrease (increase) of the frequency when increasing I0,E indicates
ING (PING), cf. Fig. 6E. These theoretical predictions are in line with
experimental evidence [2].
References
1. Viriyopase A, Memmesheimer RM, Gielen S. Cooperation and competition of gamma oscillation mechanisms. J Neurophysiol. 2016.
2. Craig MT, McBain CJ. Fast gamma oscillations are generated intrinsically
in CA1 without the involvement of fast-spiking basket cells. J Neurosci.
2015;35(8):3616–24.
O12
A discrete structure of the brain waves
Yuri Dabaghian1,2, Justin DeVito1, Luca Perotti3
1
Department of Neurology Pediatrics, Baylor College of Medicine,
Houston, TX 77030, USA; 2Department of Computational and Applied
Mathematics, Rice University, Houston, TX, 77005, USA; 3Physics
Department, Texas Southern University, 3100 Cleburne St, Houston, TX
77004, USA
Correspondence: Yuri Dabaghian ‑ dabaghian@rice.edu
BMC Neuroscience 2016, 17(Suppl 1):O12
A physiological interpretation of the biological rhythms, e.g., of the
local field potentials (LFP) depends on the mathematical and computational approaches used for its analysis. Most existing mathematical methods of the LFP studies are based on braking the signal into
a combination of simpler components, e.g., into sinusoidal harmonics
of Fourier analysis or into wavelets of the Wavelet Analysis. However, a
common feature of all these methods is that their prime components
are presumed from the onset, and the goal of the subsequent analysis
reduces to identifying the combination that best reproduces the original signal.
We propose a fundamentally new method, based on a number of deep
theorems of complex function theory, in which the prime components
of the signal are not presumed a priori, but discovered empirically [1].
Moreover, the new method is more flexible and more sensitive to the
signal’s structure than the standard Fourier method.
Applying this method reveals a fundamentally new structure in the
hippocampal LFP signals in rats in mice. In particular, our results suggest that the LFP oscillations consist of a superposition of a small,
discrete set of frequency modulated oscillatory processes, which we
call “oscillons”. Since these structures are discovered empirically, we
hypothesize that they may capture the signal’s actual physical structure, i.e., the pattern of synchronous activity in neuronal ensembles.
Proving this hypothesis will help enormously to advance a principal,
theoretical understanding of the neuronal synchronization mechanisms. We anticipate that it will reveal new information about the
structure of the LFP and other biological oscillations, which should
provide insights into the underlying physiological phenomena and
the organization of brains states that are currently poorly understood,
e.g., sleep and epilepsy.
Page 10 of 112
Acknowledgements: The work was supported by the NSF 1422438
grant and by the Houston Bioinformatics Endowment Fund.
Reference
1. Perotti L, DeVito J, Bessis D, Dabaghian Y, Dabaghian Y, Brandt VL, Frank
LM. Discrete spectra of brain rhythms (in submisison).
O13
Direction‑specific silencing of the Drosophila gaze stabilization
system
Anmo J. Kim1,†, Lisa M. Fenk1,†, Cheng Lyu1, Gaby Maimon1
1
Laboratory of Integrative Brain Function, The Rockefeller University, New
York, NY 10065, USA
Correspondence: Anmo J. Kim ‑ anmo.kim@gmail.com
†
Authors contributed equally
BMC Neuroscience 2016, 17(Suppl 1):O13
Many animals, including insects and humans, stabilize the visual
image projected onto their retina by following a rotating landscape
with their head or eyes. This stabilization reflex, also called the optomotor response, can pose a problem, however, when the animal intends
to change its gaze. To resolve this paradox, von Holst and Mittelstaedt
proposed that a copy of the motor command, or efference copy, could
be routed into the visual system to transiently silence this stabilization
reflex when an animal changes its gaze [1]. Consistent with this idea,
we recently demonstrated that a single identified neuron associated
with the optomotor response receives silencing motor-related inputs
during rapid flight turns, or saccades, in tethered, flying Drosophila [2].
Here, we expand on these results by comprehensively recording from
a group of optomotor-mediating visual neurons in the fly visual system: three horizontal system (HS) and six vertical system (VS) cells.
We found that the amplitude of motor-related inputs to each HS and
VS cell correlates strongly with the strength of each cell’s visual sensitivity to rotational motion stimuli around the primary turn axis, but
not to the other axes (Fig. 7). These results support the idea that flies
send rotation-axis-specific efference copies to the visual system during saccades—silencing the stabilization reflex only for a specific
axis, but leaving the others intact. This is important because saccades
consist of stereotyped banked turns, which involve body rotations
around all three primary axes of rotation. If the gaze stabilization
Fig. 7 The amplitudes of saccade-related potentials (SRPs) to HS
and VS cells are strongly correlated with each cell’s visual sensitivity to rightward yaw motion stimuli. A Experimental apparatus. B
Maximal-intensity z-projections of the lobula plate to visualize HS- or
VS-cell neurites that are marked by a GAL4 enhancer trap line. C, D
The amplitude of saccade-related potentials (SRPs) were inversely
correlated with visual responses, when measured under rightward
yaw motion stimuli, but not under clockwise roll motion stimuli. Each
sample point corresponds to each cell type. Error bars indicate SEM
BMC Neurosci 2016, 17(Suppl 1):54
system is impaired for only one of these axes, then the fly is expected
to attempt to maintain gaze stability, through a combination of head
and body movements, for the other two. This prediction is consistent
with behavioral measurements of head and body kinematics during
saccades in freely flying blow flies [3]. Together, these studies provide
an integrative model of how efference copies counteract a specific
aspect of visual feedback signals to tightly control the gaze stabilization system.
References
1. von Holst E, Mittelstaedt H. The principle of reafference.
Naturwissenschaften.1950;37:464–76.
2. Kim AJ, Fitzgerald JK, Maimon G. Cellular evidence for efference copy in
Drosophila visuomotor processing. Nat Neurosci. 2015;18:1247–55.
3. Schilstra C, van Hateren JH. Stabilizing gaze in flying blowflies. Nature.
1998;395:654.
O14
What does the fruit fly think about values? A model of olfactory
associative learning
Chang Zhao1, Yves Widmer2, Simon Sprecher2, Walter Senn1
1
Department of Physiology, University of Bern, Bern, 3012, Switzerland;
2
Department of Biology, University of Fribourg, Fribourg, 1700,
Switzerland
Correspondence: Chang Zhao ‑ zhao@pyl.unibe.ch
BMC Neuroscience 2016, 17(Suppl 1):O14
Associative learning in the fruit fly olfactory system has been studied
from the molecular to the behavior level [1, 2]. Fruit flies are able to
associate conditional stimuli such as odor with unconditional aversive
stimuli such as electrical shocks, or appetitive stimuli such as sugar
or water. The mushroom body in the fruit fly brain is considered to be
crucial for olfactory learning [1, 2]. The behavioral experiments show
that the learning can not be explained simply by an additive Hebbian
(i.e. correlation-based) learning rule. Instead, it depends on the timing
between the conditional and unconditional stimulus presentation. Yarali
and colleagues suggested a dynamic model on the molecular level to
explain event timing in associative learning [3]. Here, we present new
experiments together with a simple phenomenological model for learning that shows that associative olfactory learning in the fruit fly represents value learning that is incompatible with Hebbian learning.
In our model, the information of the conditional odor stimulus is conveyed by Kenyon cells from the projection neurons to the mushroom
output neurons; the information of the unconditional shock stimulus is
represented by dopaminergic neurons to the mushroom output neurons through direct or indirect pathways. The mushroom body output neurons encode the internal value (v) of the odor (o) by synaptic
weights (w) that conveys the odor information, v = w∙o. The synaptic
strength is updated according to the value learning rule, Δw = η(s − v)
õ, where s represents the (internal) strength of the shock stimulus, õ
represents the synaptic odor trace, and η is the learning rate. The
value associated with the odor determines the probability of escaping
from that odor. This simple model reproduces the behavioral data and
shows that olfactory conditioning in the fruit fly is in fact value learning. In contrast to the prediction of Hebbian learning, the escape probability for repeated odor-shock pairings is much lower than the escape
probability for a single pairing with a correspondingly stronger shock.
References
1. Aso Y, Sitaraman D, Ichinose T, Kaun KR, Vogt K, Belliart-Gurin G, Plaais
PY, Robie AA, Yamagata N, Schnaitmann C, Rowell WJ, Johnston RM,
Ngo TB, Chen N, Korff W, Nitabach MN, Heberlein U, Preat T, Branson
KM, Tanimoto H, Rubin GM: Mushroom body output neurons encode
valence and guide memory-based action selection in Drosophila. ELife.
2014;3:e04580.
2. Heisenberg M. Mushroom body memoir: from maps to models. Nat Rev
Neurosci. 2003;4:266–75.
3. Yarali A, Nehrkorn J, Tanimoto H, Herz AVM. Event timing in associative
learning: from biochemical reaction dynamics to behavioural observations. PLoS One. 2012;7(3):e32885.
Page 11 of 112
O15
Effects of ionic diffusion on power spectra of local field potentials
(LFP)
Geir Halnes1, Tuomo Mäki‑Marttunen2, Daniel Keller3, Klas H.
Pettersen4,5,Ole A. Andreassen2, Gaute T. Einevoll1,6
1
Department of Mathematical Sciences and Technology, Norwegian
University of Life Sciences, Ås, Norway; 2NORMENT, Institute of Clinical
Medicine, University of Oslo, Oslo, Norway; 3The Blue Brain Project, École
Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland;
4
Letten Centre and Glialab, Department of Molecular Medicine, Instotute
of Basic Medical Sciences, University of Oslo, Oslo, Norway; 5Centre
for Molecular Medicine Norway, University of Oslo, Oslo, Norway;
6
Department of Physics, University of Oslo, Oslo, Norway
Correspondence: Geir Halnes ‑ geir.halnes@nmbu.no
BMC Neuroscience 2016, 17(Suppl 1):O15
The local field potential (LFP) in the extracellular space (ECS) of the
brain, is a standard measure of population activity in neural tissue.
Computational models that simulate the relationship between the LFP
and its underlying neurophysiological processes are commonly used
in the interpretation such measurements. Standard methods, such as
volume conductor theory [1], assume that ionic diffusion in the ECS
has negligible impact on the LFP. This assumption could be challenged
during endured periods of intense neural signalling, under which local
ion concentrations in the ECS can change by several millimolars. Such
concentration changes are indeed often accompanied by shifts in the
ECS potential, which may be partially evoked by diffusive currents [2].
However, it is hitherto unclear whether putative diffusion-generated
potential shifts are too slow to be picked up in LFP recordings, which
typically use electrode systems with cut-off frequencies at ~0.1 Hz.
To explore possible effects of diffusion on the LFP, we developed a
hybrid simulation framework: (1) The NEURON simulator was used to
compute the ionic output currents from a small population of cortical layer-5 pyramidal neurons [3]. The neural model was tuned so that
simulations over ~100 s of biological time led to shifts in ECS concentrations by a few millimolars, similar to what has been seen in experiments [2]. (2) In parallel, a novel electrodiffusive simulation framework
[4] was used to compute the resulting dynamics of the potential and
ion concentrations in the ECS, accounting for the effect of electrical
migration as well as diffusion. To explore the relative role of diffusion,
we compared simulations where ECS diffusion was absent with simulations where ECS diffusion was included.
Our key findings were: (i) ECS diffusion shifted the local potential by
up to ~0.2 mV. (ii) The power spectral density (PSD) of the diffusionevoked potential shifts followed a 1/f2 power law. (iii) Diffusion effects
Fig. 8 Power spectrum of ECS potential in a simulation including ECS
diffusion (blue line) and a simulation without ECS diffusion (red line).
Units for frequency and power are Hz and mV2/Hz, respectively
BMC Neurosci 2016, 17(Suppl 1):54
dominated the PSD of the ECS potential for frequencies up to ~10 Hz
(Fig. 8). We conclude that for large, but physiologically realistic ECS
concentration gradients, diffusion could affect the ECS potential well
within the frequency range considered in recordings of the LFP.
References
1. Holt G, Koch C. Electrical interactions via the extracellular potential near
cell bodies. J Comput Neurosci. 1999;6:169–84.
2. Dietzel I, Heinemann U, Lux H. Relations between slow extracellular
potential changes, glial potassium buffering, and electrolyte and cellular
volume changes during neuronal hyperactivity in cat. Glia. 1989;2:25–44.
3. Hay E, Hill S, Schürmann F, Markram H, Segev I. Models of neocortical
layer 5b pyramidal cells capturing a wide range of dendritic and perisomatic active properties. PLoS Comput Biol. 2011;7(7):e1002107.
4. Halnes G, Østby I, Pettersen KH, Omholt SW, Einevoll GT: Electrodiffusive
model for astrocytic and neuronal ion concentration dynamics. PLoS
Comput Biol. 2013;9(12):e1003386.
O16
Large‑scale cortical models towards understanding relationship
between brain structure abnormalities and cognitive deficits
Yasunori Yamada1
1
IBM Research ‑ Tokyo, Japan
Correspondence: Yasunori Yamada ‑ ysnr@jp.ibm.com
BMC Neuroscience 2016, 17(Suppl 1):O16
Brain connectivity studies have revealed fundamental properties of
normal brain network organization [1]. In parallel, they have reported
structural connectivity abnormalities in brain diseases such as Alzheimer’s disease (AD) [1, 2]. However, how these structural abnormalities affect information processing and cognitive functions involved in
brain diseases is still poorly understood. To deepen our understanding
of this causal link, I developed two large-scale cortical models with
normal and abnormal structural connectivity of diffusion tensor imaging on aging APOE-4 non-carriers and carriers in the USC Multimodal
Connectivity Database [2, 3]. The possession of the APOE-4 allele is
one of the major risk factors in developing later AD, and it has known
abnormalities in structural connectivity characterized by lower network communication efficiency in terms of local interconnectivity and
balance of integration and interconnectivity [2]. The two cortical models share other parameters and consist of 2.4 million spiking neurons
and 4.8 billion synaptic connections. First, I demonstrate the biological relevance of the models by confirming that they reproduce normal
patterns of cortical spontaneous activities in terms of the following
distinctive properties observed in vivo [4]: low firing rates of individual
neurons that approximate log-normal distributions, irregular spike
trains following a Poisson distribution, a network balance between
excitation and inhibition, and greater depolarization of the average
membrane potentials. Next, to investigate how the difference in structural connectivity affects cortical information processing, I compare
cortical response properties to an input during spontaneous activity
between the cortical models. The results show that the cortical model
with the abnormal structural connectivity decreased the degree of
cortical response as well as the number of cortical regions responding to the input (Fig. 9), suggesting that the structural connectivity
abnormality observed in APOE-4 carriers might reduce cortical information propagation and lead to negative effects in information integration. Indeed, imaging studies support this suggestion by reporting
structural abnormality with lower network communication efficiency
Fig. 9 Responses to input to the left V1 in the two cortical models
with normal/abnormal structural connectivity. A Average firing rates.
B–D Cortical regions and cortical areas that significantly responded
to the input
Page 12 of 112
observed in the structural connectivity of both APOE-4 carriers and AD
patients [1, 2]. This computational approach allowing for manipulations and detailed analyses that are difficult or impossible in human
studies can help to provide a causal understanding of how cognitive
deficits in patients with brain diseases are associated with their underlying structural abnormalities.
Acknowledgements: This research was partially supported by the
Japan Science and Technology Agency (JST) under the Strategic Promotion of Innovative Research and Development Program.
References
1. Stam CJ. Modern network science of neurological disorders. Nat Rev
Neurosci. 2014;15(10):683–695.
2. Brown JA, Terashima KH, Burggren AC, Ercoli LM, Miller KJ, Small GW,
Bookheimer SY. Brain network local interconnectivity loss in aging APOE-4
allele carriers. Proc Natl Acad Sci USA. 2011;108(51):20760–5.
3. Brown JA, Rudie JD, Bandrowski A, van Horn JD, Bookheimer SY. The
UCLA multimodal connectivity database: a web-based platform for brain
connectivity matrix sharing and analysis. Front Neuroinform. 2012;6(28).
4. Ikegaya Y, Sasaki T, Ishikawa D, Honma N, Tao K, Takahashi N, Minamisawa G, Ujita S, Matsuki N. Interpyramid spike transmission stabilizes the sparseness of recurrent network activity. Cereb Cortex.
2013;23(2):293–304.
O17
Spatial coarse‑graining the brain: origin of minicolumns
Moira L. Steyn‑Ross1, D. Alistair Steyn‑Ross1
1
School of Engineering, University of Waikato, Hamilton 3240, New
Zealand
Correspondence: Moira L. Steyn‑Ross ‑ msr@waikato.ac.nz
BMC Neuroscience 2016, 17(Suppl 1):O17
The seminal experiments of Mountcastle [1] over 60 years ago established the existence of cortical minicolumns: vertical column-like
arrays of approximately 80–120 neurons aligned perpendicular to the
pial surface, penetrating all six cortical layers. Minicolumns have been
proposed as the fundamental unit for cortical organisation. Minicolumn formation is thought to rely on gene expression and thalamic
activity, but exactly why neurons cluster into columns of diameter
30–50 μm containing approximately 100 neurons is not known.
In this presentation we describe a mechanism for the formation of
minicolumns via gap-junction diffusion-mediated coupling in a network of spiking neurons. We use our recently developed method of
cortical “reblocking” (spatial coarse-graining) [2] to derive neuronal
dynamics equations at different spatial scales. We are able to show
that for sufficiently strong gap-junction coupling, there exists a minimum block size over which neural activity is expected to be coherent.
This coherence region has cross-sectional area of order (40–60 μm)2,
consistent with the areal extent of a minicolumn. Our scheme regrids
a 2D continuum of spiking neurons using a spatial rescaling theory,
established in the 1980s, that systematically eliminates high-wavenumber modes [3]. The rescaled neural equations describe the bulk
dynamics of a larger block of neurons giving “true” (rather than meanfield) population activity, encapsulating the inherent dynamics of a
continuum of spiking neurons stimulated by incoming signals from
neighbors, and buffeted by ion-channel and synaptic noise.
Our method relies on a perturbative expansion. In order for this
coarse-graining expansion to converge, we require not only a sufficiently strong level of inhibitory gap-junction coupling, but also a sufficiently large blocking ratio B. The latter condition establishes a lower
bound for the smallest “cortical block”: the smallest group of neurons
that can respond to input as a collective and cooperative unit. We
find that this minimum block-size ratio lies between 4 and 6. In order
to relate this 2D geometric result to the 3D extent of a 3-mm-thick
layered cortex, we project the cortex onto a horizontal surface and
count the number of neurons contained within each l × l grid microcell. Setting l ≈ 10 μm and assuming an average of one interneuron
per grid cell, a blocking ratio at the mid-value B = 5 implies that the
side-length of a coherent “macro-cell” will be L = Bl = 50 μm containing ~25 inhibitory plus 100 excitatory neurons (assuming an i to e
BMC Neurosci 2016, 17(Suppl 1):54
abundance ratio of 1:4) in cross-sectional area L2. Thus the minicolumn
volume will contain roughly 125 neurons. We argue that this is the
smallest diffusively-coupled population size that can support cooperative dynamics, providing a natural mechanism defining the functional
extent of a minicolumn.
We propose that minicolumns might form in the developing brain as
follows: Inhibitory neurons migrate horizontally from the ganglionic
eminence to form a dense gap-junction coupled substrate that permeates all layers of the cortex [4]. Progenitor excitatory cells ascend
vertically from the ventricular zone, migrating through the inhibitory
substrate of the cortical plate. Thalamic input provides low-level stimulus to activate spiking activity throughout the network. Inhibitory diffusive coupling allows a “coarse graining” such that neurons within a
particular areal extent respond collectively to the same input. The minimum block size prescribed by the coarse graining imposes constraints
on minicolumn geometry, leading to the spontaneous emergence of
cylindrical columns of coherent activity, each column centered on an
ascending chain of excitatory neurons and separated from neighboring chains by an annular surround of inhibition. This smallest aggregate is preferentially activated during early brain development, and
activity-based plasticity then leads to the formation of tangible structural columns.
References
1. Mountcastle VB. Modality and topographic properties of single neurons
of cat’s somatic sensory cortex. J Neurophysiol. 1957;20(4):408–34.
2. Steyn-Ross ML, Steyn-Ross DA. From individual spiking neurons to
population behavior: Systematic elimination of short-wavelength spatial
modes. Phys Rev E. 2016;93(2):022402.
3. Steyn-Ross ML, Gardiner CW. Adiabatic elimination in stochastic systems
III. Phys Rev A. 1984;29(5):2834–44.
4. Jones EG. Microcolumns in the cerebral cortex. Proc Natl Acad Sci USA.
2000;97(10):5019–21.
O18
Modeling large‑scale cortical networks with laminar structure
Jorge F. Mejias1, John D. Murray2, Henry Kennedy3, and Xiao‑Jing Wang1,4
1
Center for Neural Science, New York University, New York, NY, 10003,
USA; 2Department of Psychiatry, Yale School of Medicine, New Haven, CT,
06511, USA; 3INSERM U846, Stem Cell and Brain Research Institute, Bron
Cedex, France; 4NYU‑ECNU Institute of Brain and Cognitive Science, NYU
Shanghai, Shanghai, China
Correspondence: Jorge F. Mejias ‑ jorge.f.mejias@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):O18
Visual cortical areas in the macaque are organized according to an
anatomical hierarchy, which is defined by specific patterns of anatomical projections in the feedforward and feedback directions [1, 2].
Recent macaque studies also suggest that signals ascending through
the visual hierarchy are associated with gamma rhythms, and topdown signals with alpha/low beta rhythms [3–5]. It is not clear, however, how oscillations presumably originating at local populations can
give rise to such frequency-specific large-scale interactions in a mechanistic way, or the role that anatomical projections patterns might
have in this.
To address this question, we build a large-scale cortical network model
with laminar structure, grounding our model on a recently obtained
anatomical connectivity matrix with weighted directed inter-areal
projections and information about their laminar origin. The model
involves several spatial scales—local or intra-laminar microcircuit,
inter-laminar circuits, inter-areal interactions and large-scale cortical network—and a wide range of temporal scales—from slow alpha
oscillations to gamma rhythms. At any given level, the model is constrained anatomically and then tested against electrophysiological
observations, which provides useful information on the mechanisms
modulating the oscillatory activity at different scales. As we ascend
through the local to the inter-laminar and inter-areal levels, the model
allows us to explore the sensory-driven enhancement of gamma
rhythms, the inter-laminar phase-amplitude coupling, the relationship
between alpha waves and local inhibition, and the frequency-specific
Page 13 of 112
inter-areal interactions in the feedforward and feedback directions [3,
4], revealing a possible link with the predictive coding framework.
When we embed our modeling framework into the anatomical connectivity matrix of 30 areas (which includes novel areas not present in
previous studies [2, 6]), the model gives insight into the mechanisms
of large-scale communication across the cortex, accounts for an anatomical and functional segregation of FF and FB interactions, and predicts the emergence of functional hierarchies, which recent studies
have found in macaque [4] and human [5]. Interestingly, the functional
hierarchies observed experimentally are highly dynamic, with areas
moving across the hierarchy depending on the behavioral context
[4]. In this regard, our model provides a strong prediction: we propose
that these hierarchical jumps are triggered by laminar-specific modulations of input into cortical areas, suggesting a strong link between
hierarchy dynamics and context-dependent computations driven by
specific inputs.
References
1. Felleman DJ, Van Essen DC. Distributed hierarchical processing in the
primate cerebral cortex. Cereb Cortex. 1991;1(1):1–47.
2. Markov NT, Vezoli J, Chameau P, Falchier A, Quilodran R, Huissoud C, Lamy
C, Misery P, Giroud P, Ullman S, et al. Anatomy of hierarchy: feedforward
and feedback pathways in macaque visual cortex. J Comp Neurol.
2014;522:225–259.
3. van Kerkoerle T, Self MW, Dagnino B, Gariel-Mathis MA, Poort J, van der
Togt C, Roelfsema PR. Alpha and gamma oscillations characterize feedback and feedforward processing in monkey visual cortex. Proc Natl Acad
Sci USA. 2014;111;14332–41.
4. Bastos AM, Vezoli J, Bosman CA, Schoffelen JM, Oostenveld R, Dowdall
JR, De Weerd P, Kennedy H, Fries P. Visual areas exert feedforward and
feedback influences through distinct frequency channels. Neuron.
2015;85:390–401.
5. Michalareas G, Vezoli J, van Pelt S, Schoffelen JM, Kennedy H, Fries. Alpha–
beta and gamma rhythms subserve feedback and feedforward influences
among human visual cortical areas. Neuron. 2016;89:384–97.
6. Chaudhuri R, Knoblauch K, Gariel MA, Kennedy H, Wang XJ. A large-scale
circuit mechanism for hierarchical dynamical processing in the primate
cortex. Neuron. 2015;88:419–31.
O19
Information filtering by partial synchronous spikes in a neural
population
Alexandra Kruscha1,2, Jan Grewe3,4, Jan Benda3,4 and Benjamin Lindner1,2
1
Bernstein Center for Computational Neuroscience, Berlin, 10115,
Germany; 2Institute for Physics, Humboldt‑Universität zu Berlin, Berlin,
12489, Germany; 3Institue for Neurobiology, Eberhardt Karls Universität
Tübingen, Germany; 4Bernstein Center for Computational Neuroscience,
Munich, Germany
Correspondence: Alexandra Kruscha ‑ alexandra.kruscha@bccn‑berlin.de
BMC Neuroscience 2016, 17(Suppl 1):O19
Synchronous firing of neurons is a prominent feature in many brain
areas. Here, we are interested in the information transmission by the
synchronous spiking output of a noisy neuronal population, which
receives a common time-dependent sensory stimulus. Earlier experimental [1] and theoretical [2] work revealed that synchronous spikes
encode preferentially fast (high-frequency) components of the stimulus, i.e. synchrony can act as an information filter. In these studies a
rather strict measure of synchrony was used: the entire population has
to fire within a short time window. Here, we generalize the definition
of the synchronous output, for which only a certain fraction γ of the
population needs to be active simultaneously—a setup that seems
to be of more biological relevance. We characterize the information
transfer in dependence of this fraction and the population size, by
the spectral coherence function between the stimulus and the partial
synchronous output. We present two different analytical approaches
to derive this frequency-resolved measure (one that is more suited
for small population sizes, while the second one is applicable to
larger populations). We show that there is a critical synchrony fraction, namely the probability at which a single neuron spikes within the
BMC Neurosci 2016, 17(Suppl 1):54
predefined time window, which maximizes the information transmission of the synchronous output. At this value, the partial synchronous
output acts as a low-pass filter, whereas deviations from this critical
fraction lead to a more and more pronounced band-pass filtering
effect. We confirm our analytical findings by numerical simulations for
the leaky integrate-and-fire neuron. We also show that these findings
are supported by experimental recordungs of P-Units electroreceptors
of weakly electric fish, where the filtering effect of the synchronous
output occurs in real neurons as well.
Acknowledgement: This work was supported by Bundesministerium
für Bildung und Forschung Grant 01GQ1001A and DFG Grant 609788L1 1046/2-1.
References
1. Middleton JW, Longtin A, Benda J, Maler L. Postsynaptic receptive field
size and spike threshold determine encoding of high-frequency information via sensitivity to synchronous presynaptic activity. J Neurophysiol.
2009;101:1160–70.
2. Sharafi N, Benda J, Lindner B. Information filtering by synchronous spikes
in a neural population. J Comp Neurosc. 2013;34:285–301.
O20
Decoding context‑dependent olfactory valence in Drosophila
Laurent Badel1, Kazumi Ohta1, Yoshiko Tsuchimoto1, Hokto Kazama1
1
RIKEN Brain Science Institute, 2‑1 Hirosawa, Wako, 351‑0198, Japan
Correspondence: Laurent Badel ‑ laurent@brain.riken.jp
BMC Neuroscience 2016, 17(Suppl 1):O20
Many animals rely on olfactory cues to make perceptual decisions
and navigate the environment. In the brain, odorant molecules are
sensed by olfactory receptor neurons (ORNs), which convey olfactory
information to the central brain in the form of sequences of action
potentials. In many organisms, axons of ORNs expressing the same
olfactory receptor converge to one or a few glomeruli in the first central region (the antennal lobe in insects and the olfactory bulb in fish
and mammals) where they make contact with their postsynaptic targets. Therefore, each glomerulus can be considered as a processing
unit that relays information from a specific type of receptor. Because
different odorants recruit different sets of glomeruli, and most glomeruli respond to a wide array of odors, olfactory information at this stage
of processing is contained in spatiotemporal patterns of glomerular
activity. How these patterns are decoded by the brain to guide odorevoked behavior, however, remains largely unknown.
In Drosophila, attraction and aversion to specific odors have been
linked to the activation of one or a few glomeruli (reviewed in [1]) in
the antennal lobe (AL). These observations suggest a “labeled-line”
coding strategy, in which individual glomeruli convey signals of specific ethological relevance, and their activation triggers the execution
of hard-wired behavioral programs. However, because these studies
used few odorants, and a small fraction of glomeruli were tested, it is
unclear how the results generalize to broader odor sets, and whether
similar conclusions hold for each of the ~50 glomeruli of the fly AL.
Moreover, how compound signals from multiple glomeruli are integrated is poorly understood.
Here, we combine optical imaging, behavioral and statistical techniques to address these questions systematically. Using two-photon
imaging, we monitor Ca2+ activity in the AL in response to 84 odors.
We next screen behavioral responses to the same odorants. Comparing these data allows us to formulate a decoding model describing
how olfactory behavior is determined by glomerular activity patterns
in a quantitative manner. We find that a weighted sum of normalized glomerular responses recapitulates the observed behavior and
predicts responses to novel odors, suggesting that odor valence is
not determined solely by the activity a few privileged glomeruli. This
conclusion is supported by genetic silencing and optogenetic activation of individual ORN types, which are found to evoke modest biases
in behavior in agreement with model predictions. Finally, we test the
model prediction that the relative valence of a pair of odors depends
on the identity of other odors presented in the same experiment. We
find that the relative valence indeed changes, and may even switch,
Page 14 of 112
suggesting that perceptual decisions can be modulated by the olfactory context. Surprisingly, our model correctly captured both the
direction and the magnitude of the observed changes. These results
indicate that the valence of olfactory stimuli is decoded from AL activity by pooling contributions over a large number of glomeruli, and
highlight the ability of the olfactory system to adapt to the statistics of
its environment, similarly to the visual and auditory systems.
Reference
1. Li Q, Liberles SD. Aversion and attraction through olfaction. Curr Biol.
2015;25(3):R120–9.
P1
Neural network as a scale‑free network: the role of a hub
B. Kahng1
1
Department of Physics and Astronomy, Seoul National University, 08826,
Korea
Correspondence: B. Kahng ‑ bkahng@snu.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P1
Recently, increasing attention has been drawn to human neuroscience in network science communities. This is because recent fMRI and
anatomical experiments have revealed that neural networks of normal
human brain are scale-free networks. Thus, accumulated knowledges
in a broad range of network sciences can be naturally applied to neural networks to understand functions and properties of normal and
disordered human brain networks. Particularly, the degree exponent
value of the human neural network constructed from the fMRI data
turned out to be approximately two. This value has particularly important meaning in scale-free networks, because the number of connections to neighbors of a hub becomes largest and thus functional role
of the hub becomes extremely important. In this talk, we present the
role of the hub in pattern recognition and dynamical problems in association with neuroscience.
P2
Hemodynamic responses to emotions and decisions using
near‑infrared spectroscopy optical imaging
Nicoladie D. Tam1
1
Department of Biological Sciences, University of North Texas, Denton, TX
76203, USA
Correspondence: Nicoladie D. Tam ‑ nicoladie.tam@unt.edu
BMC Neuroscience 2016, 17(Suppl 1):P2
This study focuses on the relationship between the emotional
response, decision and the hemodynamic responses in the prefrontal cortex. This is based on the computational emotional model that
hypothesizes the emotional response is proportional to the discrepancy between the expectancy and the actuality. Previous studies had
shown that emotional responses are related to decisions [1, 2]. Specifically, the emotional responses of happy [3], sad [4], angry [5], jealous
[6] emotions are proportional to the discrepancy between what one
wants and what one gets [1, 3–7].
Methods Human subjects are asked to perform the classical behavioral economic experiment called Ultimatum Game (UG) [8]. This experimental paradigm elicits the interrelationship between decision and
emotion in human subjects [3–6]. The hemodynamic responses of the
prefrontal cortex were recorded while the subjects performed the UG
experiment.
Results The results showed that the hemodynamic response, which
corresponds to the neural activation and deactivation based on the
metabolic activities of the neural tissues, are proportional to the emotional intensity and the discrepancy between the expectancy and the
actuality. This validates the hypothesis of the proposed emotional theory [9–11] that the intensity of emotion is proportional to the disparity between the expected and the actual outcomes. These responses
are also related to the fairness perception [7], with respect to the survival functions [9, 10] similar to the responses established for happy
[1] emotion, and for fairness [12] experimentally. This is consistent with
the computational relationship between decision and fairness [13].
BMC Neurosci 2016, 17(Suppl 1):54
References
1. Tam ND. Quantification of happy emotion: dependence on decisions.
Psychol Behav Sci. 2014;3(2):68–74.
2. Tam ND. Rational decision-making process choosing fairness over monetary gain as decision criteria. Psychol Behav Sci. 2014;3(6–1):16–23.
3. Tam ND. Quantification of happy emotion: Proportionality relationship to
gain/loss. Psychol Behav Sci. 2014;3(2):60–7.
4. Tam ND: Quantitative assessment of sad emotion. Psychol Behav Sci 2015,
4(2):36-43.
5. Tam DN. Computation in emotional processing: quantitative confirmation of proportionality hypothesis for angry unhappy emotional intensity
to perceived loss. Cogn Comput. 2011;3(2):394–415.
6. Tam ND, Smith KM. Cognitive computation of jealous emotion. Psychol
Behav Sci. 2014;3(6–1):1–7.
7. Tam ND. Quantification of fairness perception by including otherregarding concerns using a relativistic fairness-equity model. Adv Soc Sci
Research J. 2014;1(4):159–69.
8. von Neumann J, Morgenstern O, Rubinstein A. Theory of games and
economic behavior. Princeton: Princeton University Press; 1953.
9. Tam D. EMOTION-I model: A biologically-based theoretical framework for
deriving emotional context of sensation in autonomous control systems.
Open Cybern Syst J. 2007;1:28–46.
10. Tam D. EMOTION-II model: a theoretical framework for happy emotion
as a self-assessment measure indicating the degree-of-fit (congruency)
between the expectancy in subjective and objective realities in autonomous control systems. Open Cybern Syst J. 2007;1:47–60.
11. Tam ND. EMOTION-III model. A theoretical framework for social empathic
emotions in autonomous control systems. Open Cybern Syst J. 2016 (in
press).
12. Tam ND: Quantification of fairness bias in relation to decisions using a relativistic fairness-equity model. Adv in Soc Sci Research J 2014, 1(4):169-178.
13. Tam ND. A decision-making phase-space model for fairness assessment.
Psychol Behav Sci. 2014;3(6–1):8–15.
P3
Phase space analysis of hemodynamic responses to intentional
movement directions using functional near‑infrared spectroscopy
(fNIRS) optical imaging technique
Nicoladie D. Tam1, Luca Pollonini2, George Zouridakis3
1
Department of Biological Sciences, University of North Texas, Denton,
TX 76203, USA; 2College of Technology, the University of Houston,
TX, 77204, USA; 3Departments of Engineering Technology, Computer
Science, and Electrical and Computer Engineering, University of Houston,
Houston, TX, 77204, USA
Correspondence: Nicoladie D. Tam ‑ nicoladie.tam@unt.edu
BMC Neuroscience 2016, 17(Suppl 1):P3
We aim to extract the intentional movement directions of the hemodynamic signals recorded from noninvasive optical imaging technique, such that a brain-computer-interface (BCI) can be built to
control a wheelchair based on the optical signals recorded from the
brain. Real-time detection of neurodynamic signals can be obtained
using functional near-infrared spectroscopy (fNIRS), which detects
both oxy-hemoglobin (oxy-Hb) and deoxy-hemoglobin (deoxy-Hb)
levels in the underlying neural tissues. In addition to the advantage of
real-time monitoring of hemodynamic signals using fNIRS over fMRI
(functional magnetic resonance imaging), fNIRS also can detect brain
signals of human subjects in motion without any movement artifacts.
Previous studies had shown that hemodynamic responses are correlated with the movement directions based on the temporal profiles
of the oxy-Hb and deoxy-Hb levels [1–5]. In this study, we will apply
a phase space analysis to the hemodynamic response to decode the
movement directions instead of using the temporal analysis in the
previous studies.
Methods In order to decode the movement directions, human subjects were asked to execute two different orthogonal directional
movements in the front-back and right-left directions while the optical hemodynamic responses were recorded in the motor cortex of the
dominant hemisphere. We aim to decode the intentional movement
directions without a priori any assumption on how arm movement
directions are correlated with the hemodynamic signals. Therefore,
Page 15 of 112
we used the phase space analysis to determine how the trajectories
of oxy-Hb and deoxy-Hb are related to each other during these arm
movements.
Results The results show that there are subpopulations of cortical
neurons that are task-related to the intentional movement directions.
Specifically, using phase space analysis of the oxy-Hb and deoxy-Hb
levels, opposite movement direction is represented by the different
hysteresis of the trajectories in opposite direction in the phase space.
Since oxy-Hb represents the oxygen delivery and deoxy-Hb represents
the oxygen extraction by the underlying brain tissues, the phase space
analysis provides a means to differentiate the movement direction by
the ratio between oxygen delivery and oxygen extraction. In other
words, the oxygen demands in the subpopulation of neurons in the
underlying tissue differ depending on the movement direction. This
also corresponds to the opposite patterns of neural activation and
deactivation during execution of opposite movement directions. Thus,
phase space analysis can be used as an analytical tool to differentiate
different movement directions based on the trajectory of the hysteresis with respect to the hemodynamic variables.
References
1. Tam ND, Zouridakis G. Optical imaging of motor cortical activation using
functional near-infrared spectroscopy. BMC Neurosci. 2012;13(Suppl
1):P27.
2. Tam ND, Zouridakis G. Optical imaging of motor cortical hemodynamic
response to directional arm movements using near-infrared spectroscopy. Int J Biol Eng. 2013;3(2):11–17.
3. Tam ND, Zouridakis G. Decoding of movement direction using optical
imaging of motor cortex. BMC Neurosci. 2013; P380.
4. Tam ND, Zouridakis G. Temporal decoupling of oxy- and deoxy-hemoglobin hemodynamic responses detected by functional near-infrared
spectroscopy (fNIRS). J Biomed Eng Med Imaging. 2014;1(2):18–28.
5. Tam ND, Zouridakis G. Decoding movement direction from motor cortex
recordings using near-infrared spectroscopy. In: Infrared spectroscopy:
theory, developments and applications. Hauppauge: Nova Science; 2014.
P4
Modeling jamming avoidance of weakly electric fish
Jaehyun Soh1, DaeEun Kim1
1
Biological Cybernetics, School of Electrical and Electronic Engineering,
Yonsei University, Shinchon, Seoul, 120‑749, South Korea
Correspondence: DaeEun Kim ‑ daeeun@yonsei.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P4
Weakly electric fish use electric field generated by the electric organ
in the tail of the fish. They detect objects by sensing the electric field
with electroreceptors on the fish’s body surface. Obstacles in the
vicinity of the fish distort the electric field generated by the fish and
the fish detect this distortion to recognize environmental situations.
Generally, weakly electric fish produce species-dependent electric
organ discharge (EOD) signals. Frequency bands of the fish’s signals
include a variety of frequencies, 50–600 Hz or higher than 800 Hz. The
EOD signals can be disturbed by similar frequency signals emitted by
neighboring weakly electric fish. They change their EOD frequencies
to avoid jamming signals when they detect the interference of signals.
This is called jamming avoidance response (JAR).
Electroreceptors of the fish read other electric fish’s EOD while they
sense their own EOD. Therefore, when two weakly electric fish are
close enough and they sense similar frequencies, their sensing ability
by EOD is impaired because of signal jamming [1, 2]. The fish lowers
its EOD frequency in response to the jamming signals when a slightly
higher frequency of signals are detected and otherwise, raises its EOD.
This response is shown in Fig. 10. The fish shift their EOD frequency
almost immediately without trial and error.
The method of how to avoid jamming has been studied for a long
time, but the corresponding neural mechanisms have not been
revealed yet so far. The JAR of Eigenmannia can be analyzed by Lissajous graphs which consist of amplitude modulations and differential
phase modulations. Relative intensity of signals at each skin can show
that the signal frequency is higher than its own signal frequency or
lower [3].
BMC Neurosci 2016, 17(Suppl 1):54
Page 16 of 112
Fig. 10 Jamming avoidance response
We suggest an algorithm of jamming avoidance for EOD signals, especially for wave-type fish. We explore the diagram of amplitude modulation versus phase modulation, and analyze the shape over the graph.
The phase differences or amplitude differences will contribute to the
estimation of the signal jamming situation. From that, the jammed
signal frequency can be detected and so it can guide the jamming
avoidance response. It can provide a special measure to predict the
jamming avoidance response. However, what type of neural structure
is available in weakly electric fish is an open question. We need further
study on this subject.
Acknowledgements: This work was supported by the National
Research Foundation of Korea (NRF) grant funded by the Korea government (MEST) (No. 2014R1A2A1A11053839).
References
1. Heiligenberg W. Electrolocation of objects in the electric fish eigenmannia
(rhamphichthyidae, gymnotoidei). J Comp Physiol. 1973;87(2):137–64.
2. Heiligenberg W. Principles of electrolocation and jamming avoidance in
electric fish. Berlin: Springer; 1977.
3. Heiligenberg W. Neural nets in electric fish. Cambridge: MIT Press; 1991.
P5
Synergy and redundancy of retinal ganglion cells in prediction
Minsu Yoo1, S. E. Palmer1,2
1
Committee on Computational Neuroscience, University of Chicago,
Chicago, IL, USA; 2Department of Organismal Biology and Anatomy,
University of Chicago, Chicago, IL, USA
Correspondence: Minsu Yoo ‑ minsu@uchicago.edu
BMC Neuroscience 2016, 17(Suppl 1):P5
Recent work has shown that retina ganglion cells (RGC) of salamanders predict future sensory information [1]. It has also been shown
that these RGC’s carry significant information about the future state
of their own population firing patterns [2]. From the perspective of
downstream neurons in the visual system that do not have independent access to the visual scene, the correlations in the RGC firing, itself,
may be important for predicting the future visual input. In this work,
we explore the structure of the generalized correlation in firing patterns in the RGC, with a particular focus on coding efficiency. From
the perspective of efficient neural coding, we might expect neurons
to code for their own future state independently (decorrelation across
cells), and to have very little predictive information extending forward
in time (decorrelation in the time domain).
In this work, we quantify whether neurons in the retina code for their
own future input independently, redundantly, or synergistically, and
how long these correlations persist in time. We use published extracellular multi-electrode data from the salamander retina in response
to repeated presentations of a natural movie [1]. We find significant
mutual information in the population firing that is almost entirely
independent except at very short time delays, where the code is
weakly redundant (Fig. 11). We also find that the information persists
to delays of up to a few 100 ms. In addition, we find that individual
neurons vary widely in the amount of predictive information they
carry about the future population firing state. This heterogeneity may
Fig. 11 Predictive information in the retinal response is coded for
independently. Red the mutual information between the binary
population firing patterns at times t and t + Δt, for 1000 randomly
selected groups of 5 cells from our 31-cell population. Time is binned
in 16.67 ms bins, and the (rare) occurrence of two spikes in a bin is
recorded as a ‘1’. Blue the sum of the mutual information between a
single cell response at time t and the future response of the group at
time t + Δt. Error bars indicate the standard error of the mean across
groups. All information quantities are corrected for finite-size effects
using quadratic extrapolation [3]
contribute to the diversity of predictive information we find across
groups in this experiment.
The results in this study may provide useful information for building a
model of the RGC population that can explain why redundant coding
is only observed at short delays, or what makes one RGC more predictive than another. Building this type of model will illustrate how the
retina represents the future.
References
1. Palmer SE, Marre O, Berry MJ, Bialek W. Predictive information in a sensory
population. Proc. Natl. Acad. Sci. 2015;112:6908–13.
2. Salisbury J, Palmer SE. Optimal prediction and natural scene statistics in
the retina. ArXiv150700125 Q-Bio [Internet]. 2015 [cited 2016 Feb 25];
Available from: http://arxiv.org/abs/1507.00125.
3. Panzeri S, Senatore R, Montemurro MA, Petersen RS. Correcting for the
sampling bias problem in spike train information measures. J. Neurophysiol. 2007;98:1064–72.
P6
A neural field model with a third dimension representing cortical
depth
Viviana Culmone1, Ingo Bojak1
1
School of Psychology, University of Reading, Reading, Berkshire, RG1 6AY, UK
Correspondence: Viviana Culmone ‑ v.culmone@pgr.reading.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P6
Neural field models (NFMs) characterize the average properties of
neural ensembles as a continuous excitable medium. So far, NFMs
have largely ignored the extension of the dendritic tree, and its influence on the neural dynamics [1]. As shown in Fig. 12A, we implement
a 3D-NFM, including the dendritic extent through the cortical layers,
starting from a well-known 2D-NFM [2]. We transform the equation
for the average membrane potential he for the point-like soma in the
2D-NFM [2] to a full cable equation form (added parts in bold):
τe
∂ 2 he (x, z, t)
∂he (x, z, t)
= − he (x, z, t) − hre + 2
∂t
∂z 2
+ f syn
ψke (he )Ike (x, z, t)
k
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 12 A The 3D-NFM adds a dendritic dimension to the 2D one
[1]. One single macrocolumn has inhibitory (I) and excitatory (E)
subpopulations. B (Top) Discretization of the dendrite. (Bottom)
Equilibrium membrane potential along the dendrite for two different
synaptic inputs. C PSDs of he for the 2D- and 3D-NFM. Increasing the
synaptic input recovers the lost alpha rhythm
The 3D-NFM is modeled considering the dendritic tree as a single
linear cable. Figure 12B shows the resulting resting potential along
the extended dendrite for synaptic input in two different locations.
Naively keeping the parameters of the 2D-NFM for the 3D-NFM results
in a power spectral density (PSD) without an alpha rhythm resonance,
see Fig. 12C. However, increasing the synaptic input by a factor fsyn can
compensate for the dispersion along the dendrite and recovers the
peak in the alpha band. We study the influence of varying the distribution of synaptic inputs along the dendritic (vertical) dimension and
of changing the (horizontal) area of the simulated cortical patch. We
also provide an outlook on how to compare our results with local field
potential recordings from real cortical tissues. We expect that 3D-NFMs
will be used widely in the future for describing such experimental
data, and that the methods used to extend the specific 2D-NFM used
here [2] will generalize to other 2D-NFMs.
References
1. Spruston N. Pyramidal neurons: dendritic structure and synaptic integration. Nat Rev Neurosci. 2008;9:206–221.
2. Bojak I, Liley DTJ. Modeling the effects of anesthesia on the electroencephalogram. Phys Rev E. 2005;71:041902.
P7
Network analysis of a probabilistic connectivity model of the
Xenopus tadpole spinal cord
Andrea Ferrario1, Robert Merrison‑Hort1, Roman Borisyuk1
1
School of Computing and Mathematics, Plymouth University, Plymouth,
PL4 8AA, United Kingdom
Correspondence: Andrea Ferrario ‑ andrea.ferrario@plymouth.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P7
Our previous results [1, 2] describe a computational anatomical model
of the Xenopus tadpole spinal cord which includes about 1400 neurons of seven types allocated on two sides of the body. This model
is based on a developmental approach, where axon growth is simulated and synapses are created (with some probability) when axons
cross dendrites. A physiological model of spiking neurons with the
Page 17 of 112
generated connectivity of about 85,000 synapses produces a very reliable swimming pattern of anti-phase oscillations in response to simulated sensory input [2].
Using the developmental model we generate 100 different sets of synaptic connections (“connectomes”), and use this information to create
a generalized probabilistic model. The probabilistic model provides
a new way to easily generate tadpole connectomes and, remarkably,
these connectomes produce similar simulated physiological behavior
to those generated using the more complex developmental approach
(e.g. they swim when stimulated). Studying these generated connectivity graphs allows us to analyze the structure of connectivity in a
typical tadpole spinal cord.
Many complex neuronal networks have been found to have “small
world” properties, including those in the nematode worm C. elegans
[3, 6], cat and macaque cortex and the human brain [4]. Small world
networks are classified between regular and random networks, and
are characterized by a high value of the clustering coefficient C and
a relatively small value of the average path length L, when compared
with Erdős-Rényi and degree matched graphs of a similar size. We used
graph theory tools to calculate the strongly connected component
of each network, which was then used to measure C and L. For the
degree-matched network, these computations have been based on
finding the probabilistic generating function [5]. By comparing these
measures with those of degree matched random graphs, we found
that tadpole’s network can be considered a small world graph. This
is also true for the sub-graph consisting only of neurons on one side
of the body, which displays properties very similar to those of the C.
elegans network. Another important subgraph, comprising only the
two main neuron types in the central pattern generator (CPG) network
also shows small world properties, but is less similar to the C. elegans
network.
Our approach allows us to study the general properties of the architecture of the tadpole spinal cord, even though in reality the actual
network varies from individual to individual (unlike in C. elegans). This
allows us to develop ideas about the organizing principles of the network, as well as to make predictions about the network’s functionality
that can be tested first in computer simulations and later in real animal experiments. In this work we combine several graph theory techniques in a novel way to analyze the structure of a complex neuronal
network where not all biological details are known. We believe that
this approach can be applied widely to analyze other animals’ nervous
systems.
References
1. Borisiuk R, al Azad AK, Conte D, Roberts A, Soffe SR. A developmental
approach to predicting neuronal connectivity from small biological datasets: a gradient-based neuron growth model. PloS One. 2014;9(2):e89461.
2. Roberts A, Conte A, Hull M, Merrison-Hort R, al Azad AK, Buhl E, Borisyuk
R, Soffe SR. Can simple rules control development of a pioneer vertebrate
neuronal network generating behavior? J Neurosci. 2014;34(2):608–21.
3. Watts DJ, Strogatz SH. Collective dynamics of ‘small-world’ networks.
Nature. 1998;440–2.
4. Kaiser M. A tutorial in connectome analysis: topological and spatial
features of brain networks. NeuroImage. 2011;892–907.
5. Newman MEJ, Strogatz SH, Watts DJ. Random graphs with arbitrary
degree distribution and their applications. Phys. Rev. 2001;E64:026118.
6. Vershney LR, Chen BL, Paniagua E, Hall DH Chklovskii DB. Structural properties of the Caenorhabditis elegans neuronal network. PloS Comput Biol.
2011;7(2):e1001066.
P8
The recognition dynamics in the brain
Chang Sub Kim1
1
Department of Physics, Chonnam National University, Gwangju, 61186,
Republic of Korea
Correspondence: Chang Sub Kim ‑ cskim@jnu.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P8
Over the years an extensive research endeavor has been given to
understanding the brain’s cognitive function in a unified principle
BMC Neurosci 2016, 17(Suppl 1):54
and to providing a formulation of the corresponding computational
scheme of the brain [1]. The explored free-energy principle (FEP)
claims that the brain’s operation on perception, learning, and action
rests on brain’s internal mechanism of trying to avoid aberrant events
encountering in its habitable environment. The theoretical measure
for this biological process has been suggested to be the informational
free-energy (IFE). The computational actualization of the FEP is carried out via the gradient descent method (GDM) in machine learning
theory.
The information content of the cognitive processes is encoded in the
biophysical matter as spatiotemporal patterns of the neuronal correlates of the external causes. Therefore, any realistic attempt to account
for the brain function must conform to the physics laws and the underlying principles. Notwithstanding the grand simplicity, however, the
FEP framework embraces some extra-physical constructs. Two major
such extra-physical constructs are the generalized motions, which are
non-Newtonian objects, and the GDM in executing the brain’s computational mechanism of perception and active inference. The GDM is
useful in finding mathematical solutions in the optimal problems, but
not derived from a physics principle.
In this work, we cast the FEP in the brain science into the framework of
the principle of least action (PLA) in physics [2]. The goal is to remove
the extra-physical constructs embedded in the FEP and to reformulate
the GDM within the standard mechanics arena. Previously, we suggested setting up the minimization scheme of the IFE in the Lagrange
mechanics formalism [3] which contained only primitive results. In the
present formulation we specify the IFE as the information-theoretic
Lagrangian and thus formally define the informational action (IA) as
time-integral of the IFE. Then, the PLA prescribes that the viable brain
minimizes the IA when encountering uninhabitable events by selecting an optimal path among all possible dynamical configurations in
the brain’s neuronal network. Specifically, the minimization yields
the mechanistic equations of motion of the brain states, which are
inverting algorithms of sensory inputs to infer their external causes.
The obtained Hamilton–Jacobi–Bellman-type equation prescribes the
brain’s recognition dynamics which do not require the extra-physical
concept of higher order motions. Finally, a neurobiological implementation of the algorithm is presented which complies with the hierarchical, operative structure of the brain. In doing so, we adopt the local
field potential and the local concentration of ions in the Hodgkin–Huxley model as the effective brain states [4]. Thus, the brain’s recognition dynamics is operatively implemented in a neuro-centric picture.
We hope that our formulation, conveying a wealth of structure as an
interpretive and mechanistic description of explaining how the brain’s
cognitive function may operate, will provide with a helpful guidance
for future simulation.
References
1. Friston K. The free-energy principle: a unified brain theory? Nat Reivew
Neurosci. 2010;11:127–38.
2. Landau LP. Classical mechanics. 2nd ed. NewYork: Springer; 1998.
3. Kim CS. The adaptive dynamics of brains: Lagrangian formulation.
Front Neurosci Conf Abstr Neuroinform. 2010. doi:10.3389/conf.
fnins.2010.13.00046.
4. Hodgkin A, Huxley A. A quantitative description of membrane current
and its application to conduction and excitation in nerve. J Physiol.
1952;117:500–44.
P9
Multivariate spike train analysis using a positive definite kernel
Taro Tezuka1
1
Faculty of Library, Information and Media Science, University of Tsukuba,
Tsukuba, 305‑0821, Japan
Correspondence: Taro Tezuka ‑ tezuka@slis.tsukuba.ac.jp
BMC Neuroscience 2016, 17(Suppl 1):P9
Multivariate spike trains, obtained by recording multiple neurons simultaneously, is a key to uncovering information representation in the brain
[1]. Other expressions used to refer to the same type of data include
“multi-neuron spike train” [2] and “parallel spike train’” [3]. One approach
Page 18 of 112
to analyze spike trains is to use kernel methods, which are known to be
among the most powerful machine learning methods. Kernel methods
rely on defining a symmetric positive-definite kernel suited to the given
data. This work proposes a general way of extending kernels on univariate (or single-unit) spike trains to multivariate spike trains.
In this work, the mixture kernel, which naturally extends a kernel
defined on univariate spike trains, is proposed and evaluated. There
are many univariate spike train kernels proposed [4–9], and the mixture kernel is applicable to any of these kernels. Considered abstractly,
a multivariate spike train is a set of time points at which different types
of events occurred. In other words, it is a sample taken from a marked
point process. The method proposed in this paper is therefore applicable to other data with the same structure.
The mixture kernel is defined as a linear combination of symmetric
positive-definite kernels on the components of the target data structure, in this case univariate spike trains. The name “mixture kernel”
derives from the common use of the word “mixture” to indicate a linear
combination in physics and machine learning, for example in Gaussian mixture models. One can prove that the mixture kernel is symmetric positive-definite if coefficient matrix of the mixture is a symmetric
positive-semidefinite matrix.
The performance of the mixture kernel was evaluated by kernel ridge
regression for estimating the value of the parameter for generating synthetic spike train data, and also the stimulus given to the animal as the
spike trains were recorded. For synthetic data, multivariate spike trains
were generated using homogenous Poisson processes. For real data,
the pvc-3 data set [2] in the CRCNS (Collaborative Research in Computational Neuroscience) data sharing website was used, which is a 10-unit
multivariate spike trains recorded from the primary visual cortex of a cat.
Acknowledgement: This work was supported in part by JSPS KAKENHI Grant Numbers 21700121, 25280110, and 25540159.
References
1. Gerstner W, Kistler WM, Naud R, Paninski L. Neuronal dynamics. Cambridge: Cambridge University Press; 2014.
2. Blanche T. Multi-neuron recordings in primary visual cortex, CRCNS.org;
2009.
3. Grun S, Rotter S. Analysis of parallel spike trains. Berlin: Springer; 2010.
4. Paiva A, Park IM, Principe JC. A reproducing kernel Hilbert space framework for spike train signal processing, Neural Comput. 2009;21(2):424–49.
5. Park IM, Seth S, Rao M, Principe JC. Strictly positive definite spike train
kernels for point process divergences. Neural Comput. 2012;24:2223–50.
6. Park IM, Seth S, Paiva A, Li L, Principe JC. Kernel methods on spike train
space for neuroscience: a tutorial. Signal Process Mag. 2013;30(4):149–60.
7. Li L, Park IM, Brockmeier AJ, Chen B, Seth S, Francis JT, Sanchez JC,
Principe JC. Adaptive inverse control of neural spatiotemporal spike
patterns with a reproducing kernel Hilbert space (RKHS) framework. IEEE
Trans Neural Syst Rehabil Eng. 2013;21(4):532–43.
8. Shpigelman L, Singer Y, Paz R, Vaadia E. Spikernels: embedding spiking neurons in inner product spaces. Adv Neural Inf Process Syst.
2003;15:125–32.
9. Eichhorn J, Tolias A, Zien A, Kuss M, Rasmussen CE, Weston J, Logothetis
N, Scholkopf B. Prediction on spike data using kernel algorithms. Adv
Neural Inf Process Syst. 2004;16:1367–74.
P10
Synchronization of burst periods may govern slow brain dynamics
during general anesthesia
Pangyu Joo1
1
Physics, POSTECH, Pohang, 37673, Republic of Korea
Correspondence: Pangyu Joo ‑ pangyu32@postech.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P10
Researchers have utilized electroencephalogram (EEG) as an important key to study brain dynamics in general anesthesia. Representative features of EEG in deep anesthesia are slow wave oscillation and
burst suppression [1], and they have so different characteristics that
they seem to have different origins. Here, we propose that the two
feature may be a different aspect of same phenomenon and show
that the slow oscillation could arise from partial synchronization of
BMC Neurosci 2016, 17(Suppl 1):54
P11
The ionic basis of heterogeneity affects stochastic synchrony
Young‑Ah Rho1,4, Shawn D. Burton2,3, G. Bard Ermentrout1,3, Jaeseung
Jeong4, Nathaniel N. Urban2,3
1
Department of Mathematics, University of Pittsburgh, Pittsburgh,
PA, USA 15260; 2Department of Biological Sciences, Carnegie Mellon
University, Pittsburgh, PA, USA 15213; 3Center for the Neural Basis
of Cognition, Pittsburgh, PA, USA 15213; 4Department of Bio and Brain
Engineering/Program of Brain and Cognitive Engineering, Korea
Advanced Institute of Science and Technology (KAIST), Daejeon, South
Korea 34141
Correspondence: Young‑Ah Rho ‑ yarho75@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P11
Synchronization in neural oscillations is a prominent feature of neural
activity and thought to play an important role in neural coding. Theoretical and experimental studies have described several mechanisms
for synchronization based on coupling strength and correlated noise
input. In the olfactory systems, recurrent and lateral inhibition mediated
by dendrodendritic mitral cell–granule cell synapses are critical for synchronization, and intrinsic biophysical heterogeneity reduce the ability
to synchronize. In our previous study, a simple phase model was used to
examine how physiological heterogeneity in biophysical properties and
firing rates across neurons affects correlation-induced synchronization
(stochastic synchrony). It has showed that heterogeneity in the firing
Introduction We estimate parameters of the inter-spike interval distributions in binaural neurons of the mammalian sound localization
neural circuit, neurons of the lateral and medial superior olive [1]. We
present equivalent descriptions of spike time probabilities using both
standard and circular statistics. We show that the difference between
sine function and beta density in the circular domain is negligible.
Results Estimation of the spike train probability density function
parameters is presented in relation to harmonic and complex sound
input. The resulting densities are expressed analytically with the use of
harmonic and Bessel functions. Parameter fits are verified by numerical simulations of spike trains (Fig. 14).
Sine
Beta density
Sine CDF
Beta CDF
Difference
A
2.5
1
1.5
1
0.5
0
ρ = 0.8
3
B
2
C
3.5
probability
References
1. Purdon PL, Pierce ET, Mukamel EA, et al. Electroencephalogram
signatures of loss and recovery of consciousness from propofol. PNAS.
2013;110(12):E1142–51.
2. Ching S, Purdon PL, Vijayan S, Kopell NJ, Brown EN. A neurophysiological–
metabolic model for burst suppression. PNAS. 2012;109(8):3095–100.
P12
Circular statistics of noise in spike trains with a periodic
component
Petr Marsalek1,2
1
Institute of Pathological Physiology, First Faculty of Medicine, Charles
University in Prague, 128 53, Czech Republic; 2Czech Technical University
in Prague, Zikova 1903/4, 166 36, Czech Republic
Correspondence: Petr Marsalek ‑ petr.marsalek@lf1.cuni.cz
BMC Neuroscience 2016, 17(Suppl 1):P12
CDF/ difference
bursting periods. To model the synchronization of burst periods,
modified version of Ching’s model of burst suppression [2] is used.
20 pyramidal neurons and 20 fast spiking neurons are divided into
10 areas composed of 2 pyramidal and 2 fast spiking neurons so that
each area exhibit burst suppression behavior independently. Then,
all the pyramidal neurons are all to all connected and the connection
strength modulates the amount of synchronization of burst periods.
The action potentials of pyramidal neurons are substituted by 1 when
the action potential larger than 0, and all other case 0. Then they are
averaged over the neurons and convoluted with 50 ms square function to see the collective activity of the neurons. As shown in Fig. 13A,
At high level of ATP recovery rate (JATP > 1), there are no suppression
period so that slow oscillation does not appear regardless of synchronization. At low level of ATP recovery rate (JATP = 0.5), we can observe
that the slow oscillation appears with increasing amplitude and finally
become burst suppression as relative connection strength increases
(Fig. 13B). When the ATP recovery rate is 0, then the pyramidal neurons
do not fire at all. These results suggest that the burst period synchronization model could explain some important features of EEG during
general anesthesia: the increasing slow oscillation amplitude as anesthesia deepen, significantly high activity in bursting period, and the
peak max phase amplitude coupling in deep anesthesia.
rates and in the shapes of the phase response curves (PRCs) reduced
output synchrony. In this study, we extend the previous phase model
to a conductance based model to examine how the density of specific
ion channels in mitral cells impacts on stochastic synchrony. A recent
study revealed that mitral cells are highly heterogeneous in the expression of the sag current, a hyperpolarization-activated inward current
(Angelo, 2011). The variability in the sag contributes to the diversity of
mitral cells and thus we wanted to know how this variability influences
synchronization. Mitral cell oscillations and bursting are also regulated
by an inactivating potassium current (IA). Based on these ion channels,
we examined the effect of changing the current densities (gA, gH) on
diversity of PRCs and of synchrony. In order to identify oscillatory patterns of bursting and repetitive spiking across gA and gH to the model,
two parameter bifurcation analysis was performed in the presence and
absence of noise. Increasing gH alone reduces the region of bursting, but
does not completely eliminate bursting, and PRCs changed much more
with respect to gA than gH. Focusing on varying gA, we next examined
a role of gA density and firing rate in stochastic synchrony by introducing the fluctuating correlated input resembling the shared presynaptic
drives. We found that heterogeneity in A-type current mainly influenced
on stochastic synchrony as we predicted in PRCs investigated theoretically, and diversity in firing rate alone didn’t account for it. In addition,
heterogeneous population with respect to gA, given decent amount
of gA density, showed better stochastic synchrony than homogeneous
population in same firing rate.
probability
Fig. 13 A The convoluted signal with different ATP recovery rates
(JATP) and relative connection strengths (C). B Standard deviation of
the convoluted signals
Page 19 of 112
0.5
ρ = 0.5
2
1.5
ρ = 0.05
1
0.5
0
0
0
0.5
phase φ
1
0
0.5
phase φ
1
0
0.5
phase φ
1
Fig. 14 Comparison of circular probability density functions of sine
and beta density. A Beta density with parameters a = b = 3.3818,
matches closely that of the sine function, used as a probability
density function (PDF). Beta density with parameters a = b = 3
solid line, is matched by sine function y = 1.05 − 1.1 cos(2π x/1.1). B
Cumulative distribution function (CDF) is shown for these densities
together with the difference between the two CDFs multiplied by
100 to visualize the comparison of the two distributions. C For testing
different vector strengths we use uniform distributions with pre-set
vector strengths (ρ = 0.8, 0.5 and 0.08)
BMC Neurosci 2016, 17(Suppl 1):54
Conclusions We use analytical techniques, where it is possible. We calculate the one-to-one correspondence of vector strength parameters
and parameters of circular distributions used for description of data.
We show here introductory figure of our paper with the two representative circular densities. We also use experimental data [2, 3] and
simulated data to compare them with these theoretical distributions.
Acknowledgements: Supported by the PRVOUK program no. 205024
at the Charles University in Prague. I acknowledge contributions to the
analytical computations by Ondrej Pokora and simulation in Matlab by
Peter G. Toth.
References
1. Bures Z, Marsalek P. On the precision of neural computation with
interaural level differences in the lateral superior olive. Brain Res.
2013;1536:16–26.
2. Joris P, Carney L, Smith P, Yin T. Enhancement of neural synchronization
in the anteroventral cochlear nucleus. I. Responses to tones at the characteristic frequency. J Neurophysiol. 1994;71(3):1022–36.
3. Joris P, Smith P, Yin T. Enhancement of neural synchronization in the
anteroventral cochlear nucleus. II. Responses in the tuning curve tail. J
Neurophysiol. 1994;71(3):1037–51.
P14
Representations of directions in EEG‑BCI using Gaussian readouts
Hoon‑Hee Kim1,2, Seok‑hyun Moon3, Do‑won Lee3, Sung‑beom Lee3,
Ji‑yong Lee3, Jaeseung Jeong1,2
1
Department of Bio and Brain Engineering and 2Program of Brain
and Cognitive Engineering, College of Engineering, Korea Advanced
Institute of Science and Technology (KAIST), Daejeon, South Korea, 34141;
3
Korea Science Academy of KAIST, Busan, South Korea, 10547
Correspondence: Jaeseung Jeong ‑ jsjeong@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P14
EEG (electroencephalography) is one of most useful neuroimaging technology and best options for BCI (Brain-Computer Interface)
because EEG has portable size, wireless and well-wearing design in
any situations. The key objective of BCI is physical control of machine
such as cursor movement in screen and robot movement [1, 2]. In previously study, the motor imagery had used for represent of direction
to movement [1, 2]. For example, the left hand imagery mapping to
move the left, the right hand imagery mapping to move the right and
both hand imagery mapping to move the forward. In this study, however, we considered only brain signals when a subject thinks directions
to movements not motor imageries. We designed the recurrent neural
networks which consist of 300–10,000 artificial linear neurons using
Echo State Networks paradigm [3]. We also recorded EEG signals using
Emotiv EPOC+ which has 16 channels (AF3, F7, F3, FC5, T7, P7, O1, O2,
P8, T8, FC6, F4, F8, AF4 and two of reference). All raw data of channels
Fig. 15 Design of recurrent neural networks and readouts
Page 20 of 112
were normalized and then used inputs to recurrent neural networks.
For representation of directions, we had built Gaussian readouts which
has preferred directions and fitted the Gaussian functions (Fig. 15). The
firing rate of readout were high when the subject thought preferred
direction. However, when the subject thought not preferred direction, the firing rate of readout slightly low down. For implement these
readouts, all of neuros in recurrent neural networks had linearly connected to all readouts and weights of these connections were trained
by linear learning rules. In result, we considered 5 healthy subjects and
recorded EEG signals for each directions. The readouts were showed
well Gaussian fitted direction preference. In this study, we considered
only two dimensions but many situations of BCI has three dimensional
space. Therefore, our study which using Gaussian readouts should be
extended to three dimensional version.
References
1. Chae Y, Jeong J, Jo S. Toward brain-actuated humanoid robots: asynchronous direct control using an EEG-based BCI. IEEE Trans Robot.
2012;28(5):1131–44.
2. LaFleur K, Cassady K, Doud A, Shades K, Rogin E, He B. Quadcopter
control in three-dimensional space using a noninvasive motor imagerybased brain–computer interface. J Neural Eng. 2013;10(4):046003.
3. Jaeger H, Haas H. Harnessing nonlinearity predicting chaotic systems and
saving energy in wireless communication. Science. 2004;304(5667):78–80.
P15
Action selection and reinforcement learning in basal ganglia
during reaching movements
Yaroslav I. Molkov1, Khaldoun Hamade2, Wondimu Teka3, William H.
Barnett1, Taegyo Kim2, Sergey Markin2, Ilya A. Rybak2
1
Department of Mathematics and Statistics, Georgia State University,
Atlanta, GA 30303, USA; 2Department of Neurobiology and Anatomy,
Drexel University, Philadelphia, PA 19129, USA; 3Department
of Mathematical Sciences, Indiana University – Purdue University,
Indianapolis, IN 46202, USA
Correspondence: Yaroslav I. Molkov ‑ ymolkov@gsu.edu
BMC Neuroscience 2016, 17(Suppl 1):P15
The basal ganglia (BG) comprise a number of interconnected nuclei
that are collectively involved in a wide range of motor and cognitive
behaviors. The commonly accepted theory is that the BG play a pivotal role in action selection and reinforcement learning facilitated by
the activity of dopaminergic neurons of substantia nigra pars compacta (SNc). These dopaminergic neurons encode prediction errors
when reward outcomes exceed or fall below anticipated values. The
BG gate appropriate behaviors from multiple moto-cortical command candidates arriving at the striatum (BG’s input nuclei) but suppress competing inappropriate behaviors. The selected motor action
is realized when the internal segment of the globus pallidus (GPi)
(BG’s output nuclei) disinhibits thalamic neurons corresponding to the
gated behavior. The BG network performs motor command selection
through the facilitation of the appropriate behavior via the “direct” striatonigral (GO) pathway and inhibition of competing behaviors by the
“indirect” striatopallidal (NOGO) pathway.
Several modeling studies have showed plausibility of the above concept in simplified cases, e.g. for binary action selection in response to
a binary cue. However, in these previous models, the possible actions/
behaviors were represented in an abstract way, and did not have a
detailed implementation as specific neuronal patterns actuating the
muscular-skeletal apparatus. To address these details, the motor system in the present study includes a 2D-biomechanical arm model in
the horizontal plane to simulate realistic reaching movements. The
arm consists of two segments (upper arm and forearm) and has two
joints (shoulder and elbow) controlled by four monoarticular (flexor
and extensor at each joint) and two bi-articular (shoulder and elbow
flexor, and shoulder and elbow extensor) muscles. The neural component of the model includes the BG, the thalamus, the motor cortex,
and spinal circuits. The low-level spinal circuitry contains six motoneurons (each controlling one muscle), and receives proprioceptor
feedback from muscles. Cortical neurons provide inputs to the spinal
BMC Neurosci 2016, 17(Suppl 1):54
network. Their activity is calculated by solving an inverse problem
(inverting the internal model) based on the initial position of the arm,
reaching distance and direction.
In the model, reaching movements in different directions were used
as a set of possible behaviors. We simulated movements in response
to a sensory cue defining the target arm position. The cortex generated signals corresponding to the cue and all possible motor commands and delivered these signals to the BG. The resulting neuronal
patterns in the motor cortex were calculated as a convolution of the
thalamic activity and all possible motor commands. The function of
BG was to establish the association between the cue and the appropriate action(s) by adjusting weights of plastic corticostriatal projections
through reinforcement learning. The BG model contained an exploratory mechanism, operating through the subthalamic nucleus (STN)
that allowed the model to constantly seek better cue-action associations that deliver larger rewards. Reinforcement learning relied on the
SNc dopaminergic signal that measured trial-to-trial changes in the
reward value, defined by performance errors.
Using this model, we simulated several learning tasks in the conditions of different unexpected perturbations. When a perturbation was
introduced, the model was capable of quickly switching away from
pre-learned associations and learning novel cue-action associations.
The analysis of the model reveals several features, that can have general importance for brain control of movements: (1) potentiation of
the cue-NOGO projections is crucial for quick destruction of preexisting cue-action associations; (2) the synaptic scaling (the decay of the
cortical-striatal synaptic weights in the absence of dopamine-mediated potentiation/depression) has a relatively short time-scale (10–20
trials); (3) quick learning is associated with a relatively poor accuracy
of the resultant movement. We suggest that BG may be involved in a
quick search for behavioral alternatives when the conditions change,
but not in the learning of skilled movements that require good
precision.
P17
Axon guidance: modeling axonal growth in T‑junction assay
Csaba Forro1, Harald Dermutz1,László Demkó1, János Vörös1
1
LBB, ETH Zürich, Zürich, 8051, Switzerland
Correspondence: Csaba Forro ‑ forro@biomed.ee.ethz.ch
BMC Neuroscience 2016, 17(Suppl 1):P17
The current field of neuroscience investigates the brain at scales varying from the whole organ, to brain slices and down to the single cell
level. The technological advances miniaturization of electrode arrays
has enabled the investigation of neural networks comprising several
neurons by recording electrical activity from every individual cell in
the network. This level of complexity is key in the study of the core
principles at play in the machinery of the brain. Indeed, it is the first
layer of complexity above the single cell that is still tractable for the
human scientist without needing to resort to a ‘Big Data’ approach. In
light of this, we strive to create topologically well-defined neural networks, akin to mathematical directed graphs, as a model systems in
order to study the basic mechanisms emerging in networks of increasing complexity and varying topology. This approach will also yield
statistically sound and reproducible observations, something which is
sought after in neuroscience [1].
The first step in realizing such a well-defined neural network is to reliably control the guidance of individual axons in order to connect the
network of cells in a controlled way. For this purpose, we present a
method consisting of obstacles forcing the axon to turn one way or
the other. The setup is made of PolyDiMethylSiloxane (PDMS) which
is microstructured by ways of state of the art photolithography procedures. Two tunnels of 5 µ height are patterned into a block of 100 µ
thick PDMS and connected in the shape of a T-junction (Fig. 16). Primary cortical neurons are inserted via entry holes at the base of the
tunnels. The entry angle of the bottom tunnel (“vertical part of the T”)
into the junction is varied between 20° (steep entry) and 90° (vertical
Page 21 of 112
Fig. 16 A The T-junction assay with an entry angle of 20°. The axon is
expected to prefer a right-turn at this angle. B A simple model is constructed where the direction of growth of the axon is proportional to
area (red) it can explore
entry). We observe that the axons prefer to turn towards the smaller
angle. We show how this observed angular selectivity in axon guidance can be explained by a simple model and how this principle
can be used to create topologically well-defined neural networks
(Fig. 16B).
Reference
1. Button KS, et al. Power failure: why small sample size undermines the
reliability of neuroscience. Nat Rev Neurosci. 2013;14(5):365–76.
P19
Transient cell assembly networks encode persistent spatial
memories
Yuri Dabaghian1,2, Andrey Babichev1,2
1
Department of Neurology Pediatrics, Baylor College of Medicine,
Houston, TX 77030, USA; 2Department of Computational and Applied
Mathematics, Rice University, Houston, TX, 77005, USA
Correspondence: Yuri Dabaghian ‑ dabaghian@rice.edu
BMC Neuroscience 2016, 17(Suppl 1):P19
The reliability of our memories is nothing short of remarkable. Thousands of neurons die every day, synaptic connections appear and
disappear, and the networks formed by these neurons constantly
change due to various forms of synaptic plasticity. How can the
brain develop a reliable representation of the world, learn and retain
memories despite, or perhaps because of, such complex dynamics?
Here we consider the specific case of spatial navigation in mammals,
which is based on mental representations of their environments—
cognitive maps—provided by the network of the hippocampal place
cells—neurons that become active only in a particular region of the
environment, known as their respective place fields. Experiments
suggest that the hippocampal map is fundamentally topological, i.e.,
more similar to a subway map than to a topographical city map, and
hence amenable to analysis by topological methods [1]. By simulating
the animal’s exploratory movements through different environments
we studied how stable topological features of space get represented
by assemblies of simulated neurons operating under a wide range of
conditions, including variations in the place cells’ firing rate, the size
of the place fields, the number of cells in the population [2,3]. In this
work, we use methods from Algebraic Topology to understand how
the dynamic connections between hippocampal place cells influence
the reliability of spatial learning. We find that although the hippocampal network is highly transient, the overall spatial map encoded by the
place cells is stable.
Acknowledgements: The work was supported by the NSF 1422438
grant and by the Houston Bioinformatics Endowment Fund.
BMC Neurosci 2016, 17(Suppl 1):54
References
1. Dabaghian Y, Brandt VL, Frank LM. Reconceiving the hippocampal map as
a topological template. eLife. 2014. doi:10.7554/eLife.03476.
2. Dabaghian Y, Mémoli F, Frank L, Carlsson G. A topological paradigm for
hippocampal spatial map formation using persistent homology. PLoS
Comput Biol. 2012;8:e1002581.
3. Arai M, Brandt V, Dabaghian Y. The effects of theta precession on spatial
learning and simplicial complex dynamics in a topological model of the
hippocampal spatial map. PLoS Comput Biol. 2014;10:e1003651.
P20
Theory of population coupling and applications to describe high
order correlations in large populations of interacting neurons
Haiping Huang1
1
RIKEN Brain Science Institute, Wako‑shi, Saitama, Japan
Correspondence: Haiping Huang ‑ physhuang@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P20
Correlations among neurons spiking activities play a prominent role
in deciphering the neural code. Various models were proposed to
understand the pairwise correlations in the population activity. Modeling these correlations sheds light on the functional organization of
the nervous system. In this study, we interpret correlations in terms of
population coupling, a concept recently proposed to understand the
multi-neuron firing patterns of the visual cortex of mouse and monkey
[1]. We generalize the population coupling to its higher order (PC2),
characterizing the relationship of pairwise firing with the population
activity. We derive the practical dimensionality reduction method for
extracting the low dimensional representation parameters, and test
our method on different types of neural data, including ganglion cells
in the salamander retina onto which a repeated natural movie was
projected [2], and layer 2/3 as well as layer 5 cortical cells in the medial
prefrontal cortex (MPC) of behaving rats [3].
For the retinal data, by considering the correlation between the pairwise firing activity and the global population activity, i.e., the second
order population coupling, the three-cell correlation could be predicted partially (64.44 %), which suggests that PC2 acts as a key circuit
variable for third order correlations. The interaction matrix revealed
here may be related to the found overlapping modular structure
of retinal neuron interactions [4]. In this structure, neurons interact
locally with their adjacent neurons, and in particular this feature is
scalable and applicable for larger networks.
About 94.79 % of three-cell correlations are explained by PC2 in the
MPC circuit. The PC2 matrix shows clear hubs’ structure in the cortical
circuit. Some neuron interacts strongly with a large portion of neurons
in the population, and such neurons may play a key role in shaping
the collective spiking behavior during the working memory task. The
hubs and non-local effects are consistent with findings reported in the
original experimental paper [3].
Acknowledgements: We are grateful to Shigeyoshi Fujisawa and
Michael J Berry for sharing us the cortical and retinal data, respectively.
We also thank Hideaki Shimazaki and Taro Toyoizumi for stimulating
discussions. This work was supported by the program for Brain Mapping by Integrated Neurotechnologies for Disease Studies (Brain/
MINDS) from Japan Agency for Medical Research and development,
AMED.
References
1. Okun M, Steinmetz NA, Cossell L, Iacaruso MF, Ko H, Bartho P, et al.
Diverse coupling of neurons to populations in sensory cortex. Nature.
2015;521:511–15.
2. Tkacik G, Marre O, Amodei D, Schneidman E, Bialek W, Berry II MJB.
Searching for collective behavior in a large network of sensory neurons.
PLoS Comput Biol. 2014;10:e1003408.
3. Fujisawa S, Amarasingham A, Harrison MT, Buzsaki G. Behavior-dependent short-term assembly dynamics in the medial prefrontal cortex. Nat
Neurosci. 2008;11:823–33.
4. Ganmor E, Segev R, Schneidman E. The architecture of functional interaction networks in the retina. J Neurosci. 2011;31(8):3044–54.
Page 22 of 112
P21
Design of biologically‑realistic simulations for motor control
Sergio Verduzco‑Flores1
1
Computational Neuroscience Unit, Okinawa Institute of Science
and Technology, Okinawa 1919‑1, Japan
Correspondence: Sergio Verduzco‑Flores ‑ sergio.verduzco@oist.jp
BMC Neuroscience 2016, 17(Suppl 1):P21
Several computational models of motor control, although apparently
feasible, fail when simulated in 3-dimensional space with redundant
manipulators [1, 2]. Moreover, it has become apparent that the details
of musculoskeletal simulations, such as the muscle model used, can
fundamentally affect the conclusions of a computational study [3].
There would be great benefits from being able to test theories involving motor control within a simulation framework that brings realism
in the musculoskeletal model, and in the networks that control movements. In particular, it would be desirable to have: (1) a musculoskeletal model considered to be research-grade within the biomechanics
community, (2) afferent information provided by standard models of
the spindle afferent and the Golgi tendon organ, (3) muscle stimulation provided by a spiking neural network that follows the basic
known properties of the spinal cord, and (4) a cerebellar network as
part of adaptive learning.
Creating this type of model is only now becoming practical, not only
due to faster computers, but due to properly validated musculoskeletal models and simulation platforms from the biomechanics community, as well as mature software and simulations techniques from the
computational neuroscience community. We show how these can be
harnessed in order to create simulations that are grounded both by
physics and by neural implementation. This pairing of computational
neuroscience and biomechanics is sure to bring further insights into
the workings of the central nervous system.
References
1. Gielen S. Review of models for the generation of multi-joint movements
in 3D. In: Sternad D, editor. Progress in motor control. New-York: Springer;
2009.
2. Verduzco-Flores SO, O’Reilly RC. How the credit assignment problems in
motor control could be solved after the cerebellum predicts increases in
error. Front Comput Neurosci. 2015;9:39.
3. Gribble PL, Ostry DJ, Sanguineti V, Laboissière R. Are complex control signals required for human arm movement? J Neurophysiol.
1998;79:1409–24.
P22
Towards understanding the functional impact of the behavioural
variability of neurons
Filipa Dos Santos1, Peter Andras1
1
School of Computing and Mathematics, Keele University,
Newcastle‑under‑Lyme, ST5 5BG, UK
Correspondence: Filipa Dos Santos ‑ f.d.s.brandao@keele.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P22
The same neuron may play different functional roles in the neural circuits to which it belongs. For example, neurons in the Tritonia pedal
ganglia may participate in variable phases of the swim motor rhythms
[1]. While such neuronal functional variability is likely to play a major
role the delivery of the functionality of neural systems, it is difficult
to study it in most nervous systems. We work on the pyloric rhythm
network of the crustacean stomatogastric ganglion (STG) [2]. Typically network models of the STG treat neurons of the same functional
type as a single model neuron (e.g. PD neurons), assuming the same
conductance parameters for these neurons and implying their synchronous firing [3, 4]. However, simultaneous recording of PD neurons
shows differences between the timings of spikes of these neurons. This
may indicate functional variability of these neurons. Here we modelled
separately the two PD neurons of the STG in a multi-neuron model of
the pyloric network. Our neuron models comply with known correlations between conductance parameters of ionic currents. Our results
reproduce the experimental finding of increasing spike time distance
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 17 The time distances between the first and second spikes of
the simulated PD neurons as a function of the gK and gCaT conductances of the neuron with variable conductances. A first spikes. B Second spikes. The PD neuron with fixed conductances had gK = 1.5768
μS and gCaT = 0.0225 μS
between spikes originating from the two model PD neurons during
their synchronised burst phase. The PD neuron with the larger calcium
conductance generates its spikes before the other PD neuron. Larger
potassium conductance values in the follower neuron imply longer
delays between spikes, see Fig. 17.
Neuromodulators change the conductance parameters of neurons
and maintain the ratios of these parameters [5]. Our results show that
such changes may shift the individual contribution of two PD neurons to the PD-phase of the pyloric rhythm altering their functionality within this rhythm. Our work paves the way towards an accessible
experimental and computational framework for the analysis of the
mechanisms and impact of functional variability of neurons within the
neural circuits to which they belong.
References
1. Hill ES, Vasireddi SK, Bruno AM, Wang J, Frost WN. Variable neuronal
participation in stereotypic motor programs. PLoS One. 2012;7:1–11.
2. Bucher D, Johnson CD, Marder E. Neuronal morphology and neuropil structure in the stomatogastric ganglion. J Comp Neurol.
2007;501:185–205.
3. Soto-Treviño C, Rabbah P, Marder E, Nadim F. Computational model of
electrically coupled, intrinsically distinct pacemaker neurons. J Neurophysiol. 2005;94:590–604.
4. Golowasch J, Casey M, Abbott LF, Marder E. Network stability from
activity-dependent regulation of neuronal conductances. Neural Comput. 1999;11:1079–96.
5. Khorkova O, Golowasch J. Neuromodulators, not activity, control coordinated expression of ionic currents. J Neurosci. 2007;27:8709–18.
Page 23 of 112
investigate gamma entrainment deficits, and at 30 Hz as a control condition. We explored the multifactoriality by performing an extensive
parameter search (approx. 4000 simulations). We focused on synaptic
and connectivity parameters of the fast spiking inhibitory interneurons in the model (i.e. number and strength of and, GABAergic decay
times at I-to-E and I-to-I connections, independently). We performed
a time–frequency analysis of simulated EEG signals and extracted the
power in the 40 Hz and the 30 Hz band, respectively. Using the power
in the 40 Hz band for 40 Hz stimulation we identified regions in the
parameter space showing strong reductions in gamma entrainment.
For these we calculated cycle-averaged EEG signals and spike time histograms of both network populations, in order to explore the dynamics underlying the reduction in gamma power.
We find three regions in the parameter space which show strong reductions in gamma power. These three regions, however, have very different parameter settings and show very different oscillatory dynamics.
The first, which produces the strongest reduction, is characterised by a
strong prolongation of decay times at I-to-E synapses and strong and
numerous I-to-E connections. Cycle-averaged spike histograms show a
broadening of distributions which indicate that the overall synchrony
is reduced, leading to the strong reduction in gamma power. However,
this parameter setting also produced a strong reduction of power in the
30 Hz control condition, which is not seen experimentally. The second
region, is characterized by prolonged I-to-I decay times together with
numerous and strong I-to-I connectivity. Here, a second peak appears in
the cycle-average spike histogram of the excitatory population, which
leads to a loss of synchrony and thus a reduction in gamma power. The
third parameter region, is also characterized by prolonged I-to-I decay
times. Moreover, it is associated with a reduction in I-to-I connection
numbers and strengths together with strong I-to-E connections. Here,
we found that in every second cycle, the spike histogram of the inhibitory neurons showed two peaks, one at the beginning and one in the
middle of the cycle. This second peak then inhibited the excitatory neurons’ response to the next stimulation. Hence, the EEG signal showed
beat-skipping, i.e. every second gamma peak was suppressed, resulting
in a decrease in gamma power.
Performing an extensive parameter search in an in silico instantiation
of an endophenotypic finding for schizophrenia, we have identified
distinct regions of the parameter space that give rise to analogous
network level behaviour found in schizophrenic patients using electrophysiology [3]. However, the oscillatory dynamics underlying this
behaviour substantially differ across regions. These regions might correspond to different subtypes of schizophrenic patients and hence,
subtypes of what might have different targets for alleviating the deficits because of their differences in underlying dynamics.
P23
Different oscillatory dynamics underlying gamma entrainment
deficits in schizophrenia
Christoph Metzner1, Achim Schweikard2, Bartosz Zurowski3
1
Science and Technology Research Institute, University of Hertfordshire,
Hatfield, United Kingdom; 2Institute for Robotics and Cognitive Systems,
University of Luebeck, Luebeck, Germany; 3Department of Psychiatry,
University of Luebeck, Schleswig–Holstein, Luebeck, Germany
Correspondence: Filipa Dos Santos ‑ c.metzner@herts.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P23
References
1. Siekmeier P. Computational modeling of psychiatric illnesses via welldefined neurophysiological and neurocognitive biomarkers. Neurosci
Biobehav Rev. 2015;57:365–80.
2. Pavão, R, Tort ABL, Amaral OB. Multifactoriality in psychiatric disorders: a computational study of schizophrenia. Schizophrenia Bull.
2015;41(4):980–88.
3. Kwon JS, O’Donnell BF, Wallenstein GV, Greene RW, Hirayasu Y, Nestor PG,
Hasselmo ME, Potts GF, Shenton ME, McCarley RW. Gamma frequencyrange abnormalities to auditory stimulation in schizophrenia. Arch Gen
Psychiatry. 1999;56(11):1001–5.
4. Beeman D. A modeling study of cortical waves in primary auditory
cortex. BMC Neurosci. 2013;14(Suppl. 1):P23.
In recent years, a significant amount of biomarkers and endophenotypic signatures of psychiatric illnesses have been identified, however, only a very limited number of computational models in support
thereof have been described so far [1]. Furthermore, the few existing
computational models typically only investigate one possible mechanism in isolation, disregarding the potential multifactoriality of the
network behaviour [2]. Here we describe a computational instantiation of an endophenotypic finding for schizophrenia, an impairment in
gamma entrainment in auditory click paradigms [3].
We used a model of primary auditory cortex from Beeman [4] and
simulated a click entrainment paradigm with stimulation at 40 Hz, to
P24
Memory recall and spike frequency adaptation
James P. Roach1, Leonard M. Sander2,3, Michal R. Zochowski2,3,4
1
Neuroscience Graduate Program, University of Michigan, Ann Arbor,
MI 48109, USA; 2Center for the Study of Complex Systems, University
of Michigan, Ann Arbor, MI 48109, USA; 3Department of Physics,
University of Michigan, Ann Arbor, MI 48109, USA; 4Biophysics Program,
University of Michigan, Ann Arbor, MI 48109, USA
Correspondence: James P. Roach ‑ roachjp@umich.edu
BMC Neuroscience 2016, 17(Suppl 1):P24
BMC Neurosci 2016, 17(Suppl 1):54
In the brain, representations of the external world are encoded by patterns of neural activity. It is critical that representations be stable, but
still easily moved between. This phenomenon has been modeled at
the network level as auto associative memory. In auto associative network models, such as the Hopfield network, representations, or memories, are stored within synaptic weights and form stable fixed points,
or attractors [1]. Spike frequency adaptation (SFA) provides a biologically plausible mechanism for switching between stabile fixed points
in the Hopfield network. In the present work we show that for low levels of SFA networks will stabilize in a representation that corresponds
to the nearest memory activity space, regardless of strength. In networks with higher levels of SFA only the pattern corresponding to the
strongest memory, or a global minimum in activity space. The effects
of SFA are similar to fast, or thermodynamic noise, but also allows for
deterministic destabilization of memories leading to periodic activation of memories through time. We argue that control of SFA level is
a universal mechanism for network-wide attractor selectivity. SFA is
tightly regulated by the neurotransmitter acetylcholine (ACh) and can
be changed on behaviorally relevant timescales. To support this claim
we demonstrate that SFA controls selectivity of spatial attractors in
a biophysical model of cholinergic modulation in cortical networks
[2, 3]. This model produces localized bumps of firing. A region with
enhanced recurrent excitation acts as an attractor for the bump location and selectivity for these regions is quickly diminishes as SFA level
increases [3]. When multiple spatial attractors of varying strengths are
stored in a network moderate increases SFA level will lead to the weak
attractors being destabilized and activity localizing within the strongest attractor. This effect is qualitatively similar to the effects of SFA in
the Hopfield network. These results indicate that ACh controls memory recall and perception within the cortex by regulation of SFA and
explain the important role cholinergic modulation plays in cognitive
functions such as attention and memory consolidation [4].
Acknowledgements: JPR was supported by an NSF Graduate
Research Fel- lowship Program under Grant No. DGE 1256260 and a
UM Rackham Merit Fellowship. MRZ and LMS were supported by NSF
PoLS 1058034.
References
1. Hopfield JJ. Neural networks and physical systems with emergent collective computational abilities. PNAS. 1982;79: 2554–8.
2. Stiefel KM, Gutkin BS, Sejnowski TJ. The effects of cholinergic neuromodulation on neuronal phase-response curves of modeled cortical neurons. J
Comp Neurosci. 2008;26:289–301.
3. Roach JP, Ben-Jacob E, Sander LM, Zochowski MR. Formation and dynamics of waves in a cortical model of cholinergic modulation. PLoS Comput
Biol. 2015;11(8): e1004449.
4. Hasselmo ME, Sarter M. Modes and models of forebrain cholinergic neuromodulation of cognition. Neuropsychopharmacology. 2011;36:52–73.
P25
Stability of neural networks and memory consolidation
preferentially occur near criticality
Quinton M. Skilling1, Nicolette Ognjanovski2, Sara J. Aton2, Michal
Zochowski1,3
1
Biophysics Program, University of Michigan, Ann Arbor, MI 48109
USA; 2Department of Molecular, Cellular, and Developmental Biology,
University of Michigan, Ann Arbor, MI, 48109 USA; 3Department
of Physics, University of Michigan, Ann Arbor, MI 48109 USA
Correspondence: Michal Zochowski ‑ michalz@umich.edu
BMC Neuroscience 2016, 17(Suppl 1):P25
Dynamic neural representations underlie cognitive processing and are
an outcome of complex interactions of network structural properties
and cellular dynamics. We have developed a new framework to study
dynamics of network representations during rapid memory formation in the hippocampus in response to contextual fear conditioning
(CFC) [1]. Experimentally, this memory paradigm is achieved by exposing mice to foot shocks while in a novel environment and later testing
for behavioral responses when reintroduced to that environment. We
employ the average minimum distance (AMD) functional connectivity
Page 24 of 112
algorithm to spiking data recorded before, during, and after CFC
using implanted stereotrodes. Comparing changes in functional connectivity using cosine similarity, we find that stable functional representations correlate well with animal performance in learning. Using
extensive computer simulations, we show that the most robust
changes compared to baseline occur when the system resides near
criticality. We attribute these results to emergence of long-range correlations during the initial process of memory formation. Furthermore,
we have developed a generic model using a generalized Hopfield
framework to link formation of novel memory representation to
functional stability changes. The network initially stores a single representation, which is to exemplify biologically already stored (old)
memories, and is then presented a new representation by freezing
a randomly chosen fraction of nodes from a novel pattern. We show
that imposing fractional input of the new representation may partially
stabilize this representation near the phase transition (critical) point.
We further show that invoking synaptic plasticity rules may fully stabilize this new representation only when the dynamics of the network
reside near criticality. Taken together these results show, for the first
time, that only when the network is at criticality can it stabilize novel
memory representations, the dynamical regime which also yields
an increase of network stability. Furthermore, our results match well
experimental data observed from CFC experiments.
Reference
1. Ognjanovski N, Maruyama D, Lashner N, Zochowski M, Aton SJ. CA1
hippocampal network activity changes during sleep-dependent memory
consolidation. Front Syst Neurosci. 2014;8:61.
P26
Stochastic oscillation in self‑organized critical states of small
systems: sensitive resting state in neural systems
Sheng‑Jun Wang1,2, Guang Ouyang2, Jing Guang3, Mingsha Zhang3, K. Y.
Michael Wong4, Changsong Zhou2,5,6
1
Department of Physics, Shaanxi Normal University, Xi’An City, ShaanXi
Province, China; 2Department of Physics and Centre for Nonlinear
Studies, Institute of Computational and Theoretical Studies, Hong Kong
Baptist University, Kowloon Tong, Hong Kong; 3State Key Laboratory
of Cognitive Neuroscience and Learning, Beijing Normal University,
Beijing, China; 4Department of Physics, Hong Kong University of Science
and Technology, Clear Water Bay, Hong Kong; 5Beijing Computational
Science Research Center, Beijing 100084, People’s Republic of China;
6
Research Centre, HKBU Institute of Research and Continuing Education,
Shenzhen, China
Correspondence: Changsong Zhou ‑ cszhou@hkbu.edu.hk
BMC Neuroscience 2016, 17(Suppl 1):P26
Self-organized critical states (SOCs) and stochastic oscillations (SOs)
are simultaneously observed in neural systems [1], which appears to
be theoretically contradictory since SOCs are characterized by scalefree avalanche sizes but oscillations indicate typical scales. Here, we
show that SOs can emerge in SOCs of small size systems due to temporal correlation between large avalanches at the finite-size cutoff,
resulting from the accumulation-release process in SOCs. In contrast,
the critical branching process without accumulation-release dynamics
cannot exhibit oscillations. The reconciliation of SOCs and SOs is demonstrated both in the sandpile model and robustly in biologically plausible neuronal networks. The oscillations can be suppressed if external
inputs eliminate the prominent slow accumulation process, providing
a potential explanation of the widely studied Berger effect or eventrelated desynchronization in neural response. The features of neural
oscillations and suppression are confirmed during task processing in
monkey eye-movement experiments. Our results suggest that finitesize, columnar neural circuits may play an important role in generating
neural oscillations around the critical states, potentially enabling functional advantages of both SOCs and oscillations for sensitive response
to transient stimuli. The results have been published in [2].
Acknowledgements: This work was partially supported by Hong
Kong Baptist University Strategic Development Fund, NSFCRGC
Joint Research Scheme HKUST/NSFC/12-13/01 (or N-HKUST 606/12),
BMC Neurosci 2016, 17(Suppl 1):54
RGC (Grants No. 604512, No. 605813, and No. 12302914), NSFC
(Grants No.11275027, No. 11328501, and No. 11305098), and the
Fundamental Research Funds for the Central Universities (Grant No.
GK201302008).
References
1. Gireesh E, Plenz D, Neuronal avalanches organize as nested theta-and
beta/gamma-oscillations during development of cortical layer 2/3. Proc
Natl Acad Sci USA. 2008;105:7576–81.
2. Wang SJ, Ouyang G, Guang J, Zhang MS, Wong KYM, Zhou CS. Stochastic
oscillation in self-organized critical states of small systems: sensitive resting state in neural systems. Phys Rev Lett. 2016;116:018101.
P27
Neurofield: a C++ library for fast simulation of 2D neural field
models
Peter A. Robinson1,2, Paula Sanz‑Leon1,2, Peter M. Drysdale1,2, Felix Fung1,2,
Romesh G. Abeysuriya3, Chris J. Rennie1,2, Xuelong Zhao1,2
1
School of Physics, University of Sydney, Sydney, New South Wales, 2006,
Australia; 2Center for Integrative Brain Function, University of Sydney,
Sydney, New South Wales, 2006, Australia
Correspondence: Paula Sanz‑Leon ‑ paula.sanz‑leon@sydney.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P27
Neural field theory [1] has addressed numerous questions regarding
brain dynamics and its interactions across many scales, becoming
a highly flexible and unified framework for the study and prediction
experimental observables of the electrical activity of the brain. These
include EEG spectra [2, 3], evoked response potentials, age-related
changes to the physiology of the brain [4], epileptic seizures [5, 6], and
synaptic plasticity phenomena [7]. However, numerical simulations of
neural field models are not widely available despite their extreme usefulness in cases where analytic solutions are less tractable. This work
introduces the features of NeuroField, a research-ready library applicable to simulate a wide range of neural field based systems involving
multiple structures (e.g., cortex, cortex and thalamic nuclei, and basal
ganglia). The link between a given neural field model, its mathematical
representation (i.e., a delay-partial differential equations system with
spatial periodic boundary conditions) and its computational implementation is described. The resulting computational model has the
capability to represent from spatially extended to neural-mass-like
systems, and it has been extensively validated against analytical solutions and against experiment [1–10]. To illustrate its flexibility, a range
of simulations modeling a variety of arousal-, sleep- and epilepsy-state
phenomena is presented [8, 9]. NeuroField has been written using
object-oriented programming in C++ and is bundled together with
MATLAB routines for quantitative offline analysis, such as spectral and
dynamic spectral analysis.
References
1. Robinson PA, Rennie CJ, Wright JJ. Propagation and stability of waves of
electrical activity in the cortex. Phys Rev E. 1997;56:826–40.
2. Robinson PA, Rennie CJ, Wright JJ, Bahramali H, Gordon E, Rowe D. Prediction of electroencephalographic spectra from neurophysiology. Phys Rev
E. 2001;63:021903.
3. Robinson PA, Rennie CJ, Rowe DL, O’Connor SC. Estimation of multiscale
neurophysiologic parameters by electroencephalographic means. Hum
Brain Mapp. 2004;23:53–72.
4. van Albada SJ, Kerr CC, Chiang AKI, Rennie CJ, Robinson PA. Neurophysiological changes with age probed by inverse modeling of EEG spectra.
Clin Neurophysiol. 2010;121:21–38.
5. Robinson PA, Rennie CJ, Rowe DL. Dynamics of large-scale brain activity in normal arousal states and epileptic seizures. Phys Rev E. 2002;
65:041924.
6. Breakspear M, Roberts JA, Terry JR, Rodrigues S, Mahant N, Robinson
PA. A unifying explanation of primary generalized seizures through
nonlinear brain modeling and bifurcation analysis. Cereb Cortex.
2006;16:1296–1313.
7. Fung PK, Haber AL, Robinson PA. Neural field theory of large-scale synaptic plasticity in the cerebral cortex. J Theor Biol. 2013; 318:44–57.
Page 25 of 112
8.
Abeysuriya RG, Rennie CJ, Robinson PA. Physiologically based arousal
state estimation and dynamics. J Neurosci Methods. 2015; 253:55–69.
9. Robinson, PA, Postnova, S, Abeysuriya, RG, Kim, JK, Roberts, JA, McKenzieSell L, Karanjai, A, Kerr, CC, Fung, F, Anderson, R, Breakspear, MJ, Drysdale,
PM, Fulcher, BD, Phillips, AKJ, Rennie, CJ, Yin G. Chapter 5: a multiscale
“working brain” model. In: Validating neurocomputational models of of
neurological and psychiatric disorders. Paris: Springer; 2015.
10. O’Connor SC, Robinson PA. Spatially uniform and nonuniform analysis of
electroencephalographic dynamics, with application to the topography
of the alpha rhythm. Phys Rev E. 2004;70:110–9.
P28
Action‑based grounding: Beyond encoding/decoding in neural
code
Yoonsuck Choe1, Huei‑Fang Yang2
1
Department of Computer Science & Engineering, Texas A&M University,
College Station, TX, 77845, USA; 2Research Center for Information
Technology Innovation, Academia Sinica, Taipei, Taiwan
Correspondence: Yoonsuck Choe ‑ choe@tamu.edu
BMC Neuroscience 2016, 17(Suppl 1):P28
How can we decode the neural activation patterns (Fig. 18A)? This is a
key question in neuroscience. We as scientists have the luxury of controlling the stimulus, based on which we can find the meaning of the
spikes (Fig. 18C-right). However, as shown in Fig. 18A (and C-left), the
problem seems intractable from the point of view of the brain itself
since neurons deeply embedded in the brain do not have direct access
to the stimulus. In [1] and related work, we showed that the decoding
problem seems intractable only because we left out the motor system
from the picture. Figure 18D shows how motor action can help processes deeply embedded in the brain can understand the meaning of
the spikes by generating motor behavior and observing the resulting
change in the neural spikes. Here, a key principle is to generate motion
that keeps the neural spike pattern invariant over time (Fig. 18E),
which allows the following to coincide (1) the property of the motion
(diagonal movement) and (2) the encoded property of the input (45°
orientation). Using reinforcement learning, we showed that the invariance criterion leads to near optimal state-action mapping for synthetic
and natural image inputs (Fig. 18F, G), where the encoded property
of the input is mapped to congruent motor action. Furthermore, we
showed that the receptive fields can be learned simultaneously with
Fig. 18 Concept (A–E) and simulation results (F–H). A Four activities without any clear meaning. b Activities in A are V1 response to
oriented lines. C Comparison of brain’s view of spikes (left; apparently
intractable) and scientist’s view of spikes (right; decoding possible).
D Visuomotor agent set up. E Invariance principle. F Ideal state(s)action(a) mapping R(s, a) (a), learned R(s, a) (b: synthetic input),
learned R(s, a) (c: natural input). G Input (a), initial gaze trajectory (b),
and learned gaze trajectory (c). H Learned state-action mapping (a:
unordered; b: reordered rows), and learned receptive fields (c: unordered; d: reordered as b) [1]
BMC Neurosci 2016, 17(Suppl 1):54
Page 26 of 112
the state-action mapping (Fig. 18H). The main lesson we learned is
that the encoding/decoding framework in neural code can lead to
a dead end unless the problem is posed from the perspective of the
brain itself; and the motor system can play an important role in the
shaping of the sensory/perceptual primitives (also see [2]).
References
1. Choe Y, Yang HF, Misra N. Motor system’s role in grounding, receptive field
development, and shape recognition. In: 7th IEEE international conference on development and learning (ICDL 2008). IEEE. p. 67–72.
2. Salinas E. How behavioral constraints may determine optimal sensory
representations. PLoS Biol. 2006;4(12):e387.
P29
Neural computation in a dynamical system with multiple time
scales
Yuanyuan Mi1,†, Xiaohan Lin1,†, Si Wu1
1
State Key Lab of Cognitive Neuroscience & Learning, IDG/McGovern
Institute for Brain Research, Beijing Normal University, Beijing 100875,
China
Correspondence: Si Wu ‑ wusi@bnu.edu.cn
†
Y.M. and X.L. contributed equally to this work
BMC Neuroscience 2016, 17(Suppl 1):P29
The brain performs computation by updating its internal states in
response to external inputs. Neurons, synapses, and the circuits are
the fundamental units for implementing brain functions. At the single neuron level, a neuron integrates synaptic inputs and generates
spikes if its membrane potential crosses the threshold. At the synapse
level, neurons interact with each other to enhance or depress their
responses. At the network level, the topology of neuronal connection pattern shapes the overall population activity. These fundamental computation units of different levels encompass rich short-term
dynamics, for example, spike-frequency adaptation (SFA) at single
neurons [1], short-term facilitation (STF) and depression (STD) at
neuronal synapses [2]. These dynamical features typically expand a
broad range of time scale and exhibit large diversity in different brain
regions. Although they play a vital part in the rise of various brain
functions, it remains unclear what is the computational benefit for
the brain to have such variability in short-term dynamics.
In this study, we propose that one benefit for having multiple dynamical features with varied time scales is that the brain can fully exploit
the advantages of these features to implement which are otherwise
contradictory computational tasks. To demonstrate this idea, we consider STF, SFA and STD with increasing time constants in the dynamics
of a CANN. The potential brain regions with these parameter values
are the sensory cortex, where the neuronal synapses are known to be
STD-dominating. We show that the network is able to implement three
seemingly contradictory computations, which are persistent activity, adaptation and anticipative tracking (see Fig. 19). Simply state, the
role of STF is to hold persistent activity in the absence of external drive,
the role of SFA is to support anticipative tracking for a moving input,
and the role of STD is to eventually suppress neural activity for a static
or transient input. Notably, the time constants of SFA and STD can be
swapped with each other, since SFA and STD have the similar effects on
the network dynamics. Nevertheless, we need to include both of them,
since a single negative feedback modulation is unable to achieve both
anticipative tracking and plateau decay concurrently. The implementation of each individual computational task based on a single dynamical
feature has been studied previously. Here, our contribution is on revealing that these tasks can be realized concurrently in a single neural circuit by combined dynamical features with coordinated time scales. We
hope that this study will shed light on our understanding of how the
brain orchestrates its rich dynamics at various levels to realize abundant cognitive functions.
Reference
1. Benda J, Herz AVM. A universal model for spike-frequency adaptation.
Neural Comput. 2003;15(11):2523–64.
2. Markram H, Wang Y, Tsodyks M. Differential signaling via the same axon of
neocortical pyramidal neurons. Proc Natl Acad Sci. 1998;95(9):5323–28.
Fig. 19 Networks implement different computations. A Persistent
activity; network can sustain activity after removing stimulus. B
Adaptation; network activity attenuates to background level given
continuous stimulus. C Anticipative tracking; D network response
leads moving stimulus in a certain speed
P30
Maximum entropy models for 3D layouts of orientation selectivity
Joscha Liedtke1,2, Manuel Schottdorf1,2, Fred Wolf1,2
1
Max Planck Institute for Dynamics and Self‑Organization, Göttingen,
Germany; 2Bernstein Center for Computational Neuroscience, Göttingen,
Germany
Correspondence: Joscha Liedtke ‑ joscha@nld.ds.mpg.de, Manuel
Schottdorf ‑ manuel@nld.ds.mpg.de
BMC Neuroscience 2016, 17(Suppl 1):P30
The neocortex is composed of 6 different layers. In the primary visual
cortex (V1), the functional architecture of basic stimulus selectivity is experimentally found to be similar across these layers [1]. The
organization in functional columns justifies the use of cortical models
describing only two-dimensional layers and disregarding functional
organization in the third dimension.
Here we show theoretically that already small deviations from an exact
columnar organization can lead to non-trivial three-dimensional functional structures (see Fig. 20). Previously, two-dimensional orientation
domains were modeled by Gaussian random fields, the maximum
entropy ensemble, allowing for an exact calculation of pinwheel densities [2]. Pinwheels are points surrounded by neurons preferring all
possible orientations and these points generalize to pinwheel strings
in three dimensions. We extend the previous two-dimensional model
characterized by its typical scale of orientation domains to a threedimensional model by keeping the typical scale in each layer and
introducing a columnar correlation length. We dissect in detail the
three-dimensional functional architecture for flat geometries and
for curved gyri-like geometries with different columnar correlation
lengths. The model is analyzed analytically complemented by numerical simulations to obtain solutions for its intrinsic statistical parameters. We find that (i) pinwheel strings are generally curved, (ii) for large
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 20 A Three-dimensional orientation domains with columnar
correlation length of Λ. B String singularities of orientation domains
in A. Typical scale of cats Λ ≈ 1 mm
curvatures closed loops and reconnecting pinwheel strings appear
and (iii) for small columnar correlation lengths a novel transition to a
rodent-like interspersed organization emerges.
This theory extends the work of [2, 3] by adding a columnar dimension
and supplements the work of [4] by a rigorous statistical treatment of
the three-dimensional functional architecture of V1. Furthermore, the
theory sheds light on the required precision of experimental techniques for probing the fine structure of the columnar organization in
V1.
References
1. Hubel DN, Wiesel TN. Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. J Physiol. 1962;160:106–54.
2. Schnabel M, Kaschube M, Löwel S, Wolf F. Random waves in the brain:
symmetries and defect generation in the visual cortex. Eur Phys J Spec
Topics. 2007;145(1):137–57.
3. Wolf F, Geisel T. Spontaneous pinwheel annihilation during visual development. Nature. 1998;395:73–8.
4. Tanaka S, Moon CH, Fukuda M, Kim SG: Three-dimensional visual feature
representation in the primary visual cortex. Neural Networks 2011,
24(10):1022-1035.
Page 27 of 112
sensory unpredictability is an effective strategy for reinforcing explorations that improve our predictive models of the world [1, 2]. However,
the computations and neural circuits underlying this unpredictabilitydependence of curiosity remain largely unknown.
A rodent model of curiosity would be useful for elucidating its underlying neural circuitry, because more specific manipulation techniques
are available than in humans. It has been shown that mice prefer
unpredictable sounds to predictable ones when the sounds are paired
with light [3]. However, frequency of stimulus presentation was a
potential confound in this study. Furthermore, a more systematic sampling of stimulus unpredictability is necessary to determine whether a
rodent analogue of the U-shaped curve indeed exists.
We have devised an operant conditioning paradigm building on [3],
using sensory stimuli as “reward” to quantify the rewardingness of
various levels of sensory predictability for rats. Rats (Long Evans, male)
are placed in a soundproofed chamber with two nosepoke holes. A
combination of sound and light stimuli is presented whenever the rat
pokes the active hole; no stimulus is associated with the inactive hole
(counterbalanced across subjects).
We hypothesize that reward is also a U-shaped function of stimulus
unpredictability in rats, and that this is due to a Bayesian precision
weighting placing more importance on deviations from reliabile predictions. This departs from previous learning-based accounts [2]. There
are five experimental conditions, systematically varied in unpredictability of the sound stimuli (as quantified by entropy H), and a control
condition, in which a nosepoke in neither hole has any consequence
(Fig. 21). Specifically, the sound stimuli are random sequences of two
possible 125-ms sound snippets of equal value to the rat, with their
frequencies of occurrence varied across conditions to vary H. Each
sequence contains eight such snippets. Across all conditions, the
light stimulus simply remains on while the sound is being played; it
is added to enhance the rats’ responding to auditory stimuli [3]. We
predict that the rats’ active nosepoke responses will be maximally
increased at intermediate H (Fig. 21).
In preliminary experiments for conditions 0 and 2 (N = 3 each; three
sessions), rats preferred the active hole to the inactive, replicating the
earlier results in mice [3]. Moreover, as hypothesized, rats responded
more to the active hole in condition 2 (mean = 15.0, SD = 5.32) than
in condition 0 (mean = 11.3, SD = 4.05); t(22) = 1.91, p = 0.0345 (onetailed t test). We note that in mice, most across-condition differences
did not emerge until around session 7 [3].
The proposed assay quantifies the rewardingness of sensory unpredictability in rats. By systematically varying the entropy of the sound
sequence, we can probe the computations behind the putative unpredictability-driven reward. The assay can furthermore be used to study
the effect of pharmacological or genetic manipulations on unpredictability-driven reward, in order to validate mechanistic implementations
of such computations.
P31
A behavioral assay for probing computations underlying curiosity
in rodents
Yoriko Yamamura1, Jeffery R. Wickens1
1
Neurobiology Research Unit, Okinawa Institute of Science
and Technology, Onna‑son, Okinawa, 904‑0412, Japan
Correspondence: Yoriko Yamamura ‑ yoriko@oist.jp
BMC Neuroscience 2016, 17(Suppl 1):P31
Curiosity in humans appears to follow an inverted U-shaped function
of unpredictability: stimuli that are neither too predictable nor too
unpredictable evoke the greatest interest [1]. Rewarding moderate
Fig. 21 Schematic of the sound stimuli used in all conditions, and
the predicted reward for each
BMC Neurosci 2016, 17(Suppl 1):54
References
1. Kidd C, Hayden BY. The psychology and neuroscience of curiosity. Neuron. 2015;88:449–60.
2. Oudeyer PY, Kaplan F. What is intrinsic motivation? A typology of computational approaches. Front Neurorobot. 2007;1:6.
3. Olsen CM, Winder DG. Stimulus dynamics increase the self-administration
of compound visual and auditory stimuli. Neurosci Lett. 2012;511:8–11.
P32
Using statistical sampling to balance error function contributions
to optimization of conductance‑based models
Timothy Rumbell1, Julia Ramsey2, Amy Reyes2, Danel Draguljić2, Patrick R.
Hof3, Jennifer Luebke4, Christina M. Weaver2
1
Computational Biology Center, IBM Research, Thomas J. Watson Research
Center, Yorktown Heights, NY 10598; 2Department of Mathematics,
Franklin and Marshall College, Lancaster, PA 17604; 3Fishberg Department
of Neuroscience and Friedman Brain Institute, Icahn School of Medicine
at Mount Sinai, New York, NY 10029; 4Department of Anatomy
and Neurobiology, Boston University School of Medicine, Boston, MA
02118
Correspondence: Christina M. Weaver ‑ christina.weaver@fandm.edu
BMC Neuroscience 2016, 17(Suppl 1):P32
Recently we developed a three-stage optimization method for fitting
conductance-based models to data [1]. The method makes novel use
of Latin hypercube sampling (LHS), a statistical space-filling design,
to determine appropriate weights automatically for various error
functions that quantify the difference between empirical target and
model output. The method uses differential evolution to fit parameters active in the subthreshold and suprathreshold regimes (below
and above action potential threshold). We have applied the method
to spatially extended models of layer 3 pyramidal neurons from the
prefrontal cortex of adult rhesus monkeys, in which in vitro action
potential firing rates are significantly higher in aged versus young
animals [2]. Here we validate our optimization method by testing its
ability to recover parameters used to generate synthetic target data.
Results from the validation fit the voltage traces of the synthetic target
data almost exactly (Fig. 22A–C), whether fitting a model with 4 ion
channels (10 parameters), or 8 ion channels (23 parameters). The optimized parameter values are either identical to, or nearby, the original
target values (Fig. 22D–F), except for a few parameters that were not
well constrained by the simulated protocols. Further, our LHS-based
scheme for weighting error functions is significantly more efficient at
recovering target parameter values than by weighting all error functions equally, or by choosing weights manually. We are now using the
method to fit models to data from several young, middle-aged, and
Fig. 22 A–C Membrane potential of the synthetic target (black), and
of randomly chosen members of the final population (colors, overlaid
almost exactly), from three validation studies. Optimized 10 and 23
parameters in A–C respectively. D–F Parameter values used to generate synthetic data (black lines), and mean ± standard deviation of
values recovered in the searches (colored circles), normalized to the
range used in the optimization
Page 28 of 112
aged monkeys. Adding new conductances to the model, and allowing altered channel kinetics in the axon initial segment versus the
soma, improves the quality of the model fits to data. We use published
results from empirical studies of layer 3 neocortical pyramidal neurons
to determine whether the optimized parameter sets are biologically
plausible.
References
1. Rumbell T, Draguljić D, Luebke J, Hof P, Weaver CM. Prediction of ion
channel parameter differences between groups of young and aged
pyramidal neurons using multi-stage compartmental model optimization. BMC Neurosci. 2015;16(Suppl. 1):P282.
2. Chang YM, Rosene DL, Killiany RJ, Mangiamele LA, Luebke JI. Increased
action potential firing rates of layer 2/3 pyramidal cells in the prefrontal
cortex are significantly related to cognitive performance in aged monkeys. Cereb Cortex. 2005;15(4):409–18.
P33
Exploration and implementation of a self‑growing
and self‑organizing neuron network building algorithm
Hu He1, Xu Yang2, Hailin Ma1, Zhiheng Xu1, Yuzhe Wang1
1
Institute of Microelectronics, Tsinghua University, Beijing, 100081, China;
2
School of Software, Beijing Institute of Technology, Beijing, 100083,
China
Correspondence: Xu Yang ‑ yangxu@tsinghua.edu.cn
BMC Neuroscience 2016, 17(Suppl 1):P33
In this work, an algorithm to build self-growing and self-organizing
neuron network according to external signals is presented, in attempt
to build neuron network with high intelligence. This algorithm takes a
bionic way to build complex neuron network. We begin with very simple external signals to provoke neurons.
In order to propagate the signals, neurons will seek to connect to each
other, thus building neuron networks. Those generated networks
will be verified and optimized, and be treated as seeds to build more
complex networks. Then we repeat this process, use more complex
external signals, and build more complex neuron networks. A parallel processing method is presented, to enhance the computation
efficiency of the presented algorithm, and to help build large scale of
neuron network with reasonable time. The result shows that, neuron
network built by our algorithm can self-grow and self-organize as the
complexity of the input external signals increase. And with the screening mechanism, neuron network that can identify different input
external signals is built successfully (Fig. 23).
Acknowledgements: This work is supported by the Core Electronic
Devices, High-End General Purpose Processor, and Fundamental
System Software of China under Grant No. 2012ZX01034-001-002,
the National Natural Science Foundation of China under Grant No.
61502032, Tsinghua National Laboratory for Information Science and
Technology (TNList), and Samsung Tsinghua Joint Laboratory.
Fig. 23 Neuron network generated by our algorithm
BMC Neurosci 2016, 17(Suppl 1):54
P34
Disrupted resting state brain network in obese subjects: a
data‑driven graph theory analysis
Kwangyeol Baek1,2, Laurel S. Morris1, Prantik Kundu3, Valerie Voon1
1
Department of Psychiatry, University of Cambridge, Cambridge, CB2
0QQ, United Kingdom; 2Department of Biomedical Engineering, Ulsan
National Institute of Science and Technology, Ulsan, South Korea;
3
Departments of Radiology and Psychiatry, Icahn School of Medicine
at Mount Sinai, New York City, 10029, USA
Correspondence: Kwangyeol Baek ‑ kb567@cam.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P34
The efficient organization and communication of brain networks
underlies cognitive processing, and disruption in resting state brain
network has been implicated in various neuropsychiatric conditions
including addiction disorder. However, few studies have focused
on whole-brain networks in the maladaptive consumption of natural rewards in obesity and binge-eating disorder (BED). Here we use
a novel multi-echo resting state functional MRI (rsfMRI) technique
along with a data-driven graph theory approach to assess global and
regional network characteristics in obesity and BED.
We collected multi-echo rsfMRI scans from 40 obese subjects (including 20 BED patients) and 40 healthy controls, and used multi-echo
independent component analysis (ME-ICA) to remove non-BOLD
noise. We estimated the normalized correlation across mean rsfMRI
Page 29 of 112
signals in 90 brain regions of the Automated Anatomical Labeling
atlas, and computed global and regional network metrics in the binarized connectivity matrix with density threshold of 5–25 %. In addition, we confirmed the observed alterations in network metrics using
the Harvard-Oxford atlas which was parcellated into 470 even-sized
regions.
Obese subjects exhibited significantly reduced global and local efficiency as well as decreased modularity in the whole-brain network
compared to healthy controls (Fig. 24). Both BED patients and the
obese subjects without BED exhibited the same alteration of network
metrics compared with healthy controls, but two obese groups did
not differ from each other. In regional network metrics, bilateral putamen, thalamus and right pallidum exhibited profoundly decreased
nodal degree and efficiency in obese subjects, and left superior frontal gyrus showed decreased nodal betweeness in obese subjects (all
p < 0.05, Bonferroni correction). Network-based statistics revealed a
cortico-striatal/cortico-thalamic network with significantly decreased
functional connectivity which consisted of bilateral putamen, pallidum, thalamus, primary motor cortex, primary somatosensory cortex,
supplementary motor area, paracentral lobule, superior parietal lobule, superior temporal cortex and left amygdala. Interestingly, when
examining the same network properties but using only single-echo
rsfMRI data analysis without ME-ICA, we find no significant differences
between groups.
Therefore, using data-driven graph theory analysis of multi-echo
rsfMRI data, we highlight more subtle impairments in cortico-striatal/
cortico-thalamic networks in obesity that have previously been associated with substance addictions. We emphasize global impairments
in network efficiency in obesity with disrupted local network organization closer to random networks. Mathematically capturing brain
network alterations in obesity provides novel insights into potential
biomarkers and therapeutic targets.
P35
Dynamics of cooperative excitatory and inhibitory plasticity
Everton J. Agnes1, Tim P. Vogels1
1
Centre for Neural Circuits and Behaviour, University of Oxford, Oxford,
OX1 3SR, UK
Correspondence: Everton J. Agnes ‑ everton.agnes@cncb.ox.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P35
Fig. 24 A Disrupted resting state brain network in obese subjects. B
Global network properties network-based statistics
Neurons receive balanced excitatory and inhibitory inputs, a phenomenon thought to be essential for a variety of computations [1–3].
Inhibitory synaptic plasticity is an obvious candidate for imposing this
balanced input regime [2,4], leaving excitatory synapses available to
learn patterns and memories. Recent experimental work seems to
agree with that notion of collaborative excitatory and inhibitory plasticity [4], but recent models do not take direct interactions into consideration. Instead, learning rules are usually tuned to indirectly but
constructively interact via the firing-rates they elicit [3,5]. Without
proper parameter tuning, this can be problematic because excitatory
and inhibitory synaptic plasticity models may have different homeostatic set points, making synaptic weights fluctuate wildly (Fig. 25A,
B; green lines). Here we present a hybrid model of inhibitory synaptic
plasticity that combines the simplicity of spike-based models with
the addition of a excitatory/inhibitory input dependence. It captures
recent experimental findings showing that changes at inhibitory synapses are strongly correlated with the balance between excitation and
inhibition and that inhibitory synapses do not change when excitatory input is blocked [4]. Essentially, our model is a symmetric spiketiming-dependent plasticity (STDP) rule in which the learning-rate is
controlled by excitatory and inhibitory activities—a spike-timing- and
current-dependent plasticity (STCDP) model. Balance is maintained,
but the learning rule does not impose fixed-point attractor dynamics to post-synaptic neurons, because there is no change in inhibitory synapses once the total input is balanced. Inhibitory synapses
change depending on excitatory synapses, which means that plasticity depends on at least three synaptic participants (trisynaptic) instead
of only two (bisynaptic). We show that when combined with an excitatory synaptic plasticity model, both excitatory and inhibitory weights
converge to stable values, as the firing-rate reaches the fixed-point
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 25 A Schematics representing the neuronal network. A
group of 2000 excitatory neurons and 500 inhibitory neurons are
recurrently connected with sparse connectivity and the excitatory
neurons receive random input from an external pool of neurons. B
Excitatory neurons’ mean firing-rate (top), mean excitatory weight
onto excitatory neurons (middle) and mean inhibitory weight onto
excitatory neurons (connections marked as plastic in A). Simulation
of the neuronal network with a spike-based inhibitory learning rule
is represented by green lines (STDP) while simulation with our novel
spike-timing- and current-dependent learning rule is shown in yellow
(STCDP). The dashed lines represent the fixed points imposed by the
excitatory (high) and inhibitory (low) learning rules. The low fixed
point only exists for the inhibitory STDP model (simulation represented by the green lines)
imposed by the excitatory learning rule (Fig. 25B; yellow lines). More
importantly, the learning rule allows efficient and stable learning of
new weights when the balance is disrupted, opening the door for
effective and stable learning of arbitrary synaptic patterns.
Acknowledgements: This work was partially funded by the Brazilian
agency CNPq (Grant Agreement Number 235144/2014-2) and the Sir
Henry Dale Fellowship (Grant Agreement WT100000).
References
1. Denève S, Machens CK. Efficient codes and balanced networks. Nat
Neurosci. 2016;19:375–85.
2. Vogels TP, Froemke RC, Doyon N, Gilson M, Haas JS, Liu R, Maffei A, Miller
P, Wierenga CJ, Woodin MA, et al. Inhibitory synaptic plasticity: spike
timing-dependence and putative network function. Front Neural Circuits.
2013;7:119.
3. Zenke F, Agnes EJ, Gerstner W. Diverse synaptic plasticity mechanisms
orchestrated to form and retrieve memories in spiking neural networks.
Nat Commun. 2015;6:6922.
4. D’amour JA, Froemke RC. Inhibitory and excitatory spike-timing-dependent plasticity in the auditory cortex. Neuron. 2015;86:514–28.
5. Sprekeler H, Clopath C, Vogels TP. Interactions of excitatory and inhibitory
synaptic plasticity. Front Comp.
P36
Frequency‑dependent oscillatory signal gating in feed‑forward
networks of integrate‑and‑fire neurons
William F. Podlaski1, Tim P Vogels1
1
Centre for Neural Circuits and Behaviour, University of Oxford, Oxford, UK
Correspondence: William F. Podlaski ‑ william.podlaski@cncb.ox.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P36
Neural oscillations—the periodic synchronisation of neuronal spiking—is a common feature of brain activity, with several hypothesised
functions relating to information flow, attention and brain state [1].
Previous experimental work has shown that oscillatory activity correlates with moments of heightened attention, and that communication
between different brain areas is often marked by an increase in oscillatory coherence between the regions [2]. Theoretical and modelling
work has helped to explore the mechanisms behind neuronal oscillations, and some of their effects on neural coding and signal propagation [3]. Recently, theoretical studies have explored how resonance
Page 30 of 112
might affect signal processing [4, 5] and how information can be propagated along different pathways according to oscillatory phase and
frequency [6].
We expand this work here by studying how resonance at the single neuron level might be used for frequency-dependent gating of information flow in neuronal networks. We show that in feed-forward spiking
network simulations background oscillations can synchronise or desynchronise the spikes of a propagated signal, changing its content and
emphasis from rate code to synfire code or vice versa. Such a mechanism
can modulate information flow without rewiring the signal pathways
themselves, allowing to select for specific downstream readout targets.
Building on this idea, we can create entire pathways that can be selectively (in-)activated by different background oscillatory frequencies
without changing the connectivity of the network. We hypothesise that
neuronal resonance, combined with resonance in synapses and network
motifs, can allow for precise oscillatory gating of information in cortex.
Building on previous studies of resonance and oscillatory signal propagation [4,5,6] we propose a plausible mechanism for how fast and precise
frequency-dependent gating might be achieved in the brain.
Acknowledgements: Research was supported by a Sir Henry Dale
Royal Society and Wellcome Trust Research Fellowship (WT100000).
References
1. Buzsáki G. Rhythms of the brain. Oxford: Oxford University Press; 2011.
2. Engel AK, Fries P, Singer W. Dynamic predictions: oscillations and synchrony in top-down processing. Nat Rev Neurosci. 2001;2(10):704–16.
3. Wang XJ. Neurophysiological and computational principles of cortical
rhythms in cognition. Physiol Rev. 2010;90:1195–1268.
4. Richardson MJE, Brunel N, Hakim V. From subthreshold to firing-rate
resonance. J Neurophysiol. 2003;89:2538–54.
5. Hahn G, Bujan AF, Frégnac Y, Aertsen A, Kumar A. Communication
through resonance in spiking neuronal networks. PLoS Comput Biol.
2014;10(8):e1003811.
6. Akam T, Kullmann DM. Oscillatory multiplexing of population codes for
selective communication in the mammalian brain. Nat Rev Neurosci.
2014;15(2):111–22.
P37
Phenomenological neural model for adaptation of neurons
in area IT
Martin Giese1, Pradeep Kuravi2, Rufin Vogels2
1
Section Computational Sensomotorics, CIN & HIH, Department
of Cognitive Neurology, University Clinic Tübingen, Germany; 2Lab. Neuro
en Psychofysiologie, Dept. Neuroscience, KU Leuven, Belgium
Correspondence: Martin Giese ‑ martin.giese@uni‑tuebingen.de
BMC Neuroscience 2016, 17(Suppl 1):P37
For repeated stimulation neurons in higher-level visual cortex show
adaptation effects. Such effects likely influence repetition suppression paradigms in fMRI studies and the formation of high-level aftereffects, e.g. for faces [1]. A variety of theoretical explanations has been
discussed, which are difficult to distinguish without detailed electrophysiological data [2]. Meanwhile, detailed physiological experiments
on the adaptation of shape-selective neurons in inferotemporal cortex
(area IT) have provided constraints that help to narrow down possible
neural processes. We propose a neurodynamical model that reproduces a number of these experimental observations by biophysically
plausible neural circuits. Our model uses the mean-field limit and consists of a neural field of shape-selective dynamic linear-threshold neurons that are augmented several adaptation processes: (i) spike-rate
adaptation; (ii) an input fatigue adaptation process, modeling adaptation in earlier hierarchy levels and of afferent synapses; (iii) a firing-rate
fatigue adaptation process that models adaptation dependent on the
output firing rates of the neurons. The model with a common parameter set is compared to results from several studies about adaptation in
area IT. The model reproduces the following experimentally observed
effects: (i) shape of the typical PSTHs of IT neurons; (ii) temporal
decay for repeated stimulation of the same neurons with many repetitions of the same stimulus [3] (Fig. 26A); (iii) dependence of adaptation on efficient and ineffective adaptor stimuli, which stimulate the
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 26 Simulation results. A Decay of neural activity for multiple
repetitions of the same stimulus. B Experiment adapting with effective and ineffective stimuli. C Dependence of the PSTH on adaptor
duration and unadapted response (black)
neuron strongly or only moderately [4] (Fig. 26B); (iv) dependence of
the strength of the adaptation effect on the duration of the adaptor
(Fig. 26C). A mean field model with several additional adaptive processes can account for the observed experimental effects, where all
introduced processes were necessary to account for the results. Especially the observed dependence on the effectivity of the adaptor cannot be reproduced without an appropriate mixture if an input fatigue
and a firing-rate fatigue mechanism. This suggests that adaptation in
IT neurons is significantly influenced by several biophysical processes
with different spatial and temporal scales.
Acknowledgements: Supported by EC Fp7-PEOPLE-2011-ITN
PITN-GA-011-290011 (ABC), FP7-ICT-2013-FET-F/604102 (HBP), FP7ICT-2013-10/611909 (Koroibot), BMBF, FKZ: 01GQ1002A, DFG GI
305/4-1 + KA 1258/15-1.
References
1. Leopold DA, O’Toole AJ, Vetter T, Blanz V: Prototype-referenced
shape encoding revealed by high-level aftereffects. Nat Neurosci.
2001;4(1):89–94.
2. Grill-Spector K, Henson R, Martin A. Repetition and the brain: neural
models of stimulus-specific effects. Trends Cogn Sci. 2006;10(1):14–23.
3. Sawamura H, Orban GA, Vogels R. Selectivity of neuronal adaptation does
not match response selectivity: a single-cell study of the FMRI adaptation
paradigm. Neuron. 2006;49(2):307–18.
4. De Baene W, Vogels R. Effects of adaptation on the stimulus selectivity
of macaque inferior temporal spiking activity and local field potentials.
Cereb Cortex. 2010;20(9):2145–65.
P38
ICGenealogy: towards a common topology of neuronal ion
channel function and genealogy in model and experiment
Alexander Seeholzer1,†, William Podlaski2,†, Rajnish Ranjan3, Tim Vogels2
1
Laboratory of Computational Neuroscience, EPF Lausanne, Switzerland;
2
Centre for Neural Circuits and Behaviour, University of Oxford, UK; 3The
Blue Brain Project, EPF Lausanne, Switzerland
Correspondence: Alexander Seeholzer ‑ alex.seeholzer@epfl.ch
†
These authors contributed equally to this work.
BMC Neuroscience 2016, 17(Suppl 1):P38
Ion channels are fundamental constituents determining the function
of single neurons and neuronal circuits. To understand their complex
interactions, the field of computational modeling has proven essential:
since its emergence, thousands of ion channel models have been created and published as part of detailed neuronal simulations [1]. Faced
with this large variety of models, it is difficult to determine how particular models relate to each other, to the interpretability of simulations and, importantly, to experimental data.
Here, we present a framework within which we analyzed a pilot set of
2378 voltage- or calcium-dependent published ion channel models
for the NEURON simulator [1]. We extracted annotated metadata from
Page 31 of 112
Fig. 27 A Visualizations available on the web-resource [2] for model
browsing. B Schematic of upload and evaluation. Both experimental
current traces and mod files can be uploaded to our servers, where
they are scored and compared to all models currently in the database. C Exemplary result of automated comparison: Current traces
(recorded from “Ramp” and “Activation” voltage clamp protocols) of
the uploaded model (red) together with mean (1st, 2nd, 3rd, 4th) and
individual (gray) traces of the four most similar clusters of channel
models in the database
all associated publications, helping identify their use in simulations
(e.g. the animal type, neuron type or area of compartmental models)
and the provenance of ion channel models as they were derived from
other published work. This categorical and relational metadata is combined with quantitative evaluations of all channel models: individual
channels are characterized by their responses to voltage clamp protocols. With subsequent cluster analysis, we extract topologies of ion
channel similarity and genealogy, identifying redundancy and groups
of common channel kinetics.
The result of this large-scale assay of published work is freely accessible through interactive visualizations (see Fig. 27A) on the Ion Channel
Genealogy (ICG) web-resource [2], providing a tool for model discovery
and comparison. Bridging the gap between model and experiment,
our resource allows classifying new channel models and experimental
current traces within the topology of all models currently in the database (see Fig. 27B, C). The ICG framework thus allows for quantitative
comparison of ion channel kinetics, experimental and model alike,
aimed to facilitate field-wide standardization of experimentally-constrained modeling.
Acknowledgements: Research was supported by a Sir Henry Dale
Royal Society & Wellcome Trust Research Fellowship (WT100000).
A.S. was supported by the Swiss National Science Foundation
(200020_147200). R.R. was supported by the EPFL Blue Brain Project
Fund and the ETH Board funding to the Blue Brain Project.
References
1. Hines ML, Morse T, Migliore M, Carnevale NT, Shepherd GM. ModelDB. A
database to support computational neuroscience. J Comput Neurosci.
2004;17:7–11.
2. ICGenealogy Project Website. http://icg.neurotheory.ox.ac.uk.
P39
Temporal input discrimination from the interaction
between dynamic synapses and neural subthreshold oscillations
Joaquin J. Torres1, Fabiano Baroni2, Roberto Latorre3, Pablo Varona3
1
Departamento de Electromagnetismo y Física de la Materia,
and Institute “Carlos I” for Theoretical and Computational Physics,
University of Granada, Granada, Spain; 2School of Psychological Sciences,
Faculty of Biomedical and Psychological Sciences, Monash University,
Australia; 3Grupo de Neurocomputación Biológica, Dpto. de Ingeniería
Informática, Escuela Politécnica Superior, Universidad Autónoma de
Madrid, Spain
Correspondence: Pablo Varona ‑ pablo.varona@uam.es
BMC Neuroscience 2016, 17(Suppl 1):P39
Neuronal subthreshold oscillations underlie key mechanisms of information discrimination in single cells while dynamic synapses provide
channel-specific input modulation. Previous studies have shown
that intrinsic neuronal properties, in particular subthreshold oscillations, constitute a biophysical mechanism for the emergence of nontrivial single-cell input/output preferences (e.g., preference towards
BMC Neurosci 2016, 17(Suppl 1):54
decelerating vs. accelerating input trains of the same average rate) [1,
2]. It has also been shown that short-term synaptic dynamics, in the
form of short-term depression and/or short-term facilitation, can provide a channel-specific mechanism for the enhancement of the postsynaptic effects of temporally specific input sequences [3, 4]. While
intrinsic oscillations and synaptic dynamics are typically studied independently, it is reasonable to hypothesize that their interplay can lead
to more selective and complex temporal input processing.
Here, we extend and refine our previous computational study on the
interaction between subthreshold oscillations and synaptic depression [5]. In particular, we investigated whether, and under which
conditions, the combination of intrinsic subthreshold oscillations
and short-term synaptic dynamics can act synergistically to enable
the emergence of robust and channel-specific selectivity in neuronal
input–output transformations. We calculated analytically the voltage
trajectories and spike output of generalized integrate-and-fire (GIF)
model neurons in response to temporally distinct trains of input EPSPs.
In particular, we considered triplets of input EPSPs in a range that covers intrinsic and synaptic time scales, and analyzed the model output
as intrinsic and synaptic parameters were varied.
Our results show that intrinsic and synaptic dynamics interact in a
complex manner for the emergence of specific input–output transformations. In particular, precise non-trivial preferences emerge from
synergistic intrinsic and synaptic preferences, while broader selectivity is observed for mismatched intrinsic and synaptic dynamics. We
discuss the conditions for robustness of the observed input/output
relationships.
We conclude that the interaction of intrinsic and synaptic properties
can enable the biophysical implementation of complex and channel-specific mechanisms for the emergence of selective neuronal
responses. We further interpret our results in the light of experimental
evidence describing distinct short-term synaptic dynamics in different
afferents converging onto the same neuron, as in the case of parallel
and climbing fiber inputs to cerebellar Purkinje cells, and advance specific hypotheses that link heterogeneous synaptic dynamics of distinct
pathways onto the same post-synaptic target to their distinct computational function. We also discuss the impact of single-channel/singleneuron temporal input discrimination in the context of information
processing based on heterogeneous elements.
Page 32 of 112
Neural activity in awake primate early visual cortex exhibits transients
with intervals of 250-300 ms. Experimental work by us and others has
shown that these transients are related to microsaccadic eye movements [1, 2]. These short transients are followed by periods of steady
activity that last until the next microsaccade (Fig. 28A).
We found that computational models of excitatory-inhibitory spiking networks organized in a structure of columns and hypercolumns,
are able to represent relevant stimulus information when subjected
to 3–4 Hz saccade-like transients. The simulated networks expressed
evoked responses with power in the alpha–beta band (~8–25 Hz) as
well as gamma rhythmic activity (~25–80 Hz) similar to in vivo local
field recordings in monkey V1 (Fig. 28A, B).
We show that in phase I, the model produces large-scale spatial synchrony and pronounced alpha–beta power. In phase II the model
exhibits narrow-band gamma oscillations with spatially local synchrony. The activity in the model network (rate and timing coding) in
phase I mainly reflects feedforward input (Fig. 28C, D), whereas, the
network activity in phase II was dominated by recurrent connections
(Fig. 28C, E).
The model network activity closely matches that found in experiments. The simulation results suggest that transient phase (phase I)
allows for resetting the network and rapid feedfoward processing of
novel information, whereas detailed processing and contextualization by recurrent activity take place in the period of steady gamma
activity (phase II). Therefore we arrived at hypotheses on the functional interpretation of phases I and II that can be possibly tested in
Acknowledgements: We acknowledge support from MINECO
FIS2013-43201-P, DPI2015-65833-P, TIN-2012-30883 and ONRG Grant
N62909-14-1-N279.
References
1. Baroni F, Varona P. Subthreshold oscillations and neuronal input–output
relationships. Neurocomputing. 2007;70:1611–14.
2. Baroni F, Torres JJ, Varona P. History-dependent excitability as a single-cell
substrate of transient memory for information discrimination. PLoS One.
2010;5:e15023.
3. O’Donnell C, Nolan MF. Tuning of synaptic responses: An organizing principle for optimization of neural circuits. Trends Neurosci. 2011;34:51–60.
4. Torres JJ, Kappen HJ. Emerging phenomena in neural networks with
dynamic synapses and their computational implications. Front Comp
Neurosci. 2013;7.
5. Latorre R, Torres JJ, Varona P. Interplay between subthreshold oscillations
and depressing synapses in single neurons. PLoS One. 2016;11:e0145830.
P40
Different roles for transient and sustained activity during active
visual processing
Bart Gips1,†, Eric Lowet1,2,†, Mark J Roberts1,2, Peter de Weerd2, Ole Jensen1,
Jan van der Eerden1
1
Radboud University, Donders Institute for Brain, Cognition
and Behaviour, 6525 EN Nijmegen, The Netherlands; 2Faculty
of Psychology and Neuroscience, Maastricht University, 6200 MD
Maastricht, the Netherlands
Correspondence: Bart Gips ‑ bart.gips@donders.ru.nl
† Authors have made equal contribution
BMC Neuroscience 2016, 17(Suppl 1):P40
Fig. 28 A Time–frequency representation of local field potential
(LFP) locked to a microsaccade (MS) recorded in primate V1. B Time–
frequency representation of simulated LFP. C Schematic representation of the model network illustrating input (injection current),
recurrent connection pattern and output (spike trains). D The input
to the neurons is best reflected in the simulated spike trains (output)
during phase I, quantified by mutual information (MI). E Recurrent
connection pattern is best reflected in the output during phase II
BMC Neurosci 2016, 17(Suppl 1):54
an experimental setup. First, because of the reset of network activity
by a microsaccade, phase I is the optimal time window to switch information flow among competing networks through a top-down signal.
This indicates that signals related to visual attention are most likely to
occur just after a saccade. Second, the increased efficacy of recurrent
connections during phase II indicate that contextualization operations
such as figure-ground segregation [3] and contour completion occur
in the steady phase ~100 ms after the onset of a (micro)saccade.
References
1. Lowet E, Roberts MJ, Bosman CA, Fries P, de Weerd P. Areas V1 and V2
show microsaccade-related 3–4 Hz covariation in gamma power and
frequency. Eur J Neurosci. 2015.
2. Martinez-Conde S, Otero-Millan J, Macknik SL. The impact of microsaccades on vision: towards a unified theory of saccadic function. Nat Rev
Neurosci. 2013;14:83–96.
3. Self MW, van Kerkoerle T, Supèr H, Roelfsema PR. Distinct roles of the
cortical layers of area V1 in figure-ground segregation. Curr Biol. 2013:1–9.
Page 33 of 112
advantages like fasciliation of alteration of functional patterns, optimization of information transfer and maximization of correlation length,
shows striking robustness against structural deficits. Taking into
account brain’s long-range anatomical connections and compensatory
mechanisms like neuroplasticity, if the results of this study are generalizable to the brain, they may help to explain the delay in clinical diagnosis of multiple neurodegenerative diseases in which possible deficit
in functional connectivity among brain regions contribute to the cognitive dysfunctions.
References
1. Chialvo DR. Emergent complex neural dynamics. Nat Phys. 2010;6:744–50.
2. Eguíluz VM, Chialvo DR, Cecchi GA, Baliki M, Apkarian AV. Scale-free brain
functional networks. Phys Rev Let.t 2005;94:018102.
3. Beggs J, Plenz D. Neuronal avalanches in neocortical circuits. J Neurosci.
2003;23:11167–77.
4. Fraiman D, Balenzuela P, Foss J, Chialvo D. Ising-like dynamics in largescale functional brain networks. Phys Rev E. 2009;79:061922.
5. Zarepour M, Niry MD, Valizadeh A. Functional scale-free networks in the
two-dimensional Abelian sandpile model. Phys Rev E. 2015;92:012822.
P41
Scale‑free functional networks of 2D Ising model are highly
robust against structural defects: neuroscience implications
Abdorreza Goodarzinick1, Mohammad D. Niry1,2, Alireza Valizadeh1,3
1
Department of Physics, Institute for Advanced Studies in Basic Sciences
(IASBS), Zanjan 45137‑66731, Iran; 2Center for Research in Climate Change
and Global Warming (CRCC), Institute for Advanced Studies in Basic
Sciences (IASBS), Zanjan 45137‑66731, Iran; 3School of Cognitive Sciences,
Institute for Research in Fundamental Sciences (IPM), Tehran ‑ Iran
Correspondence: Abdorreza Goodarzinick ‑ a.goodarzinick@iasbs.ac.ir
BMC Neuroscience 2016, 17(Suppl 1):P41
P42
High frequency neuron can facilitate propagation of signal
in neural networks
Aref Pariz1, Shervin S Parsi1, Alireza Valizadeh1,2
1
Department of Physics, Institute for advanced studies in basic sciences,
Zanjan, Iran; 2School of Cognitive Sciences, Institute for Studies
in Theoretical Physics and Mathematics, Niavaran, Tehran, Iran
Correspondence: Aref Pariz ‑ a.pariz@iasbs.ac.ir
BMC Neuroscience 2016, 17(Suppl 1):P42
In recent years, several experimental observations have confirmed
the emergence of self-organized criticality (SOC) in the brain at different scales [1]. At large scale, functional brain networks obtained
from fMRI data have shown that node-degree distributions and probability of finding a link versus distance are indicative of scale-free and
small-world networks regardless of the tasks in which the subjects
were involved [2]. At small scale, the study of neuronal avalanches in
networks of living neurons revealed power-law behavior in both spatial and temporal scales [3]. It is also shown that functional networks
of the brain are strikingly similar to those derived from the 2D Ising
model at critical temperature [4] and the 2D abelian sandpile model
[5].
The importance to see whether brain network’s scaling properties
associated with healthy conditions are altered under various pathologies and how structural defects of a system at criticality can affect its
functional connectivity motivated us to study robustness of functional networks of 2D Ising model at critical point against elimination of structural sites. The results showed that the statistics of the
functional network indicative of criticality (evident in healthy brain
controls), such as power-law behavior and small-worldness remained
robust against random elimination of structural sites up to percolation limit (see Fig. 29). The resulting functional network maintained
its key properties orders of magnitude higher than those of the same
system poised in a super-critical or sub-critical state. These results
can show that self-organized critical behavior, besides having unique
Signal transmission is of interest from both fundamental and clinical
perspective and has been well studied in nonlinear science and complex networks [1, 2]. In particular, in nervous systems, cognitive processing involves signal propagation through multiple brain regions
and the activation of large numbers of specific neurons [3–6]. In
information propagation through brain regions, each part, known as
generator, activated locally as information comes to it from neighboring generators. Although the problem is well studied in the context
of complex networks, our focus here is on the effect of the intrinsic
dynamical properties of the reciprocal generators on the propagation
of signal.
In this study we explored the propagation of information in a chain of
neurons and networks. As signal propagate through the chain of networks, the firing rate of networks show a fluctuation as host network
(the network which receive signal). Here the response is the amplitude
of fast Fourier transform of firing rates of each network. If the host network has sufficiently higher intrinsic firing rate than others, signal can
transfer with higher amplitude, otherwise, other networks will not get
affected. As a result of propagation of signal, for the former case, all
networks will show a peak in frequency domain at exactly the same
frequency as input signal (Fig. 30A), but with different amplitude
which show the efficacy of transmitted information. Also the same
result can obtain by a chain of single LIF neurons (Fig. 30B). As phase
response curve of the chain and it response to signal show, if the host
neuron has higher firing rate (call it leader neuron), the propagation
Fig. 29 Relevant parameters of functional network of 2D Ising
model at critical point versus fraction of defect to the structural cells.
A Power-law exponent of degree-distribution, B small-worldness
measure, C average degree
Fig. 30 Inhomogeneity of input current on host network, increases
the response of network. A, B Response of networks of neurons and
chain of neurons, for different inhomogeneity on host network and
host neuron, respectively
BMC Neurosci 2016, 17(Suppl 1):54
of information will be enhanced. But this higher firing rate has a limit
which after that the whole chain will act asynchronously and results
the loss of information was aimed to propagate.
References
1. Liang X, Liu Z, Li B. Weak signal transmission in complex networks and its
application in detecting connectivity. Phys Rev E. 2010;80:046102.
2. Perc M. Stochastic resonance on weakly paced scale-free networks. Phys
Rev E. 2008;78:036105.
3. Abeles M. Corticonics: neural circuits of the cerebral cortex. Cambridge:
Cambridge UP; 1991.
4. Aertsen A, Diesmann M, Gewaltig MO. Propagation of synchronous spiking activity in feedforward neural networks. J Physiol. 1996;90:243–247.
5. van Rossum MC, Turrigiano GG, Nelson SB. Fast propagation of firing rates
through layered networks of noisy neurons. J Neurosci. 2002;22:1956–66.
6. Vogels TP, Abbott LF. Signal propagation and logic gating in networks of
integrate-and-fire neurons. J Neurosci. 2005;25(46):10786–95.
P43
Investigating the effect of Alzheimer’s disease related
amyloidopathy on gamma oscillations in the CA1
region of the hippocampus
Julia M. Warburton1, Lucia Marucci2, Francesco Tamagnini3,4, Jon Brown3,4,
Krasimira Tsaneva‑Atanasova5
1
Bristol Centre for Complexity Sciences, University of Bristol, Bristol, BS8
1TR, UK; 2Department of Engineering Mathematics, University of Bristol,
Bristol, BS8 1UB, UK; 3School of Physiology and Pharmacology, University
of Bristol, Bristol, BS8 1TD, UK; 4Medical School, University of Exeter, Exeter,
EX4 4PE, UK; 5Department of Mathematics, University of Exeter, Exeter,
EX4 4QF, UK
Correspondence: Julia M. Warburton ‑ julia.warburton@bristol.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P43
Alzheimer’s disease (AD) is the main form of dementia and is characterised clinically by cognitive decline and impairments to memory
function. One of the key histopathological features of AD thought to
cause this neurodegeneration is the abnormal aggregation of the protein amyloid-β (Aβ) [1]. Transgenic mouse models that overexpress Aβ
are used to investigate the potential functional consequences of this
amyloidopathy in AD. In this study we use in vitro electrophysiology
data recorded from PDAPP transgenic mice (a mouse model of amyloidopathy) and their wild-type littermates to parameterise a hippocampal network model [2]. The aim of the study is to investigate
how amyloidopathy alters gamma frequency oscillations within the
hippocampus, which is one of the regions first affected in AD.
We use a synaptically connected network of excitatory pyramidal neurons and inhibitory interneurons to simulate the gamma frequency
activity [3]. Each cell is described by a single-compartment Hodgkin–Huxley type equation, with the properties of the voltage-gated
channels fit to the intrinsic properties measured experimentally,
which included stimulated firing frequency data and the associated
action potentials from CA1 pyramidal neurons and three-types of CA1
interneuron. Network activity is either driven deterministically via a
direct stimulus, such as a step pulse or a theta wave, or via a stochastic
input. We perform power spectral density analysis to analyse the oscillatory activity.
Our model focuses on gamma frequency oscillations, which lie in the
30–100 Hz range, because of the associations with attention, sensory processing and potentially of most relevance to AD, with learning and memory. It has been shown that within the hippocampus
gamma oscillations enable cross-talk between distributed cell assemblies, with low frequency gamma associated with coupling between
the CA1 and the CA3 region and fast frequency gamma associated
with coupling between the CA1 and the medial entorhinal cortex [4].
EEG measurements from AD mouse models have identified network
hypersynchrony alongside decreased gamma activity, with the role
of interneurons in this process highlighted. [5]. By incorporating the
pyramidal neuron and interneuron data in our model we aim to learn
more about which parameters are most significant in these effects and
Page 34 of 112
to further understanding of the effects of amyloidopathy on oscillatory activity.
Acknowledgements: This work was supported by funding from the
EPSRC.
References
1. Hardy J, Selkoe DJ. The amyloid hypothesis of Alzheimer’s disease: progress and problems on the road to therapeutics. Science. 2002;297:353–6.
2. Kerrigan TL, Brown JT, Randall TL. Characterization of altered intrinsic
excitability in hippocampal CA1 pyramidal cells of the Aβ-overproducing
PDAPP mouse. Neuropharmacology. 2014;79:515–24.
3. Kopell NJ, Borgers C, Pervouchine D, Maerba P, Tort A. Gamma and theta
rhythms in biophysical models of hippocampal circuits. In: Hippocampal
microcircuits: a computational modeler’s resource book, chap 15; p.
423–57.
4. Colgin LL, Denninger T, Fyhn M, Hafting T, Bonnevie T, Jensen O, Moser
M-B, Moser EI. Frequency of gamma oscillations routes flow of information in the hippocampus. Nature. 2009;462:353–7.
5. Verret L, et al. Inhibitory interneuron deficit links altered network activity
and cognitive dysfunction in Alzheimer model. Cell. 2012;149:708–21.
P44
Long‑tailed distributions of inhibitory and excitatory weights in a
balanced network with eSTDP and iSTDP
Florence I. Kleberg1, Jochen Triesch1
1
Frankfurt Institute for Advanced Studies, Frankfurt am Main, Hessen,
Germany, 60438
Correspondence: Florence I. Kleberg ‑ kleberg@fias.uni‑frankfurt.de
BMC Neuroscience 2016, 17(Suppl 1):P44
The strengths of excitatory synapses in cortex and hippocampus have
been shown to follow a rightward-skewed or long-tailed distribution [1,2]. Such distributions can be achieved in recurrent balanced
networks [3, 4], after synaptic modification by spike-timing dependent plasticity (STDP) [5] and synaptic scaling [6]. Recently, long-tailed
distributions have also been observed for inhibitory synapses in cultured cortical neurons [7], confirming early findings in hippocampal
slices [8]. However, the conditions and plasticity mechanisms necessary for achieving long-tailed distributions of inhibitory synapses are
unknown. Furthermore, different forms of inhibitory STDP have been
reported, but their effect on the distribution of inhibitory synaptic efficacies are largely unknown [9-11].
Here we investigate how plasticity in the inhibitory synapses in a selforganised recurrent neural network (SORN [12]) with leaky integrateand-fire neurons can lead to long-tailed distributions of synaptic
weights. We examine different inhibitory STDP (iSTDP) rules and characterize the conditions under which right-skewed shapes of inhibitory
synaptic weight distributions are obtained while a balance between
excitation and inhibition is maintained. While the ratio of long-term
potentiation to long-term depression in iSTDP affects the shape
of the distribution, a variety of window shapes for iSTDP can each
achieve long-tailed distributions of inhibitory weights. We find that
a precise balance of excitation and inhibition can be achieved with a
strongly right-skewed distribution of inhibitory weights. Our results
suggest that long-tailed distributions of inhibitory weights could be
a ubiquitous feature of neural circuits that employ different plasticity
mechanism.
References
1. Bekkers, JM, Stevens, CF. NMDA and non-NMDA receptors are colocalized at individual excitatory synapses in cultured rat hippocampus.
Nature. 1989;341:230–3.
2. Loewenstein Y, Kuras A, Rumpel S. Multiplicative dynamics underlie the
emergence of the log-normal distribution of spine sizes in the neocortex
in vivo. J Neurosci. 2011;31(26):9481–8.
3. Effenberger F, Jost J, Levina A. Self-organization in balanced state
networks by STDP and homeostatic plasticity. PLoS Comput Biol.
2015;11(9):e1004420.
BMC Neurosci 2016, 17(Suppl 1):54
Miner D, Triesch J. Plasticity-driven self-organization under topological
constraints accounts for non-random features of cortical synaptic wiring.
PLoS Comput Biol. 2016;12(2):e1004759.
5. Bi G, Poo M. Synaptic modifications in cultured hippocampal neurons:
dependence on spike timing, synaptic strength, and postsynaptic cell
type. J Neurosci. 1998;18(24):10464–72.
6. Turrigiano GG, Leslie KR, Desai NS, Rutherford LC, Nelson SB. Activitydependent scaling of quantal amplitude in neocortical neurons. Nature.
1998;391(6670):892–6.
7. Rubinski A, Ziv NE. Remodeling and tenacity of inhibitory synapses:
relationships with network activity and neighboring excitatory synapses.
PLoS Comput Biol. 2015;11(11):e1004632.
8. Miles R. Variation in strength of inhibitory synapses in the CA3 region of
guinea-pig hippocampus in vitro. J Physiol. 1990;431:659–76.
9. Woodin MA, Ganguly K, Poo M. Coincident pre-and postsynaptic activity
modifies GABAergic synapses by postsynaptic changes in Cl− transporter activity. Neuron. 2003;39(5):807–20.
10. Haas JS, Nowotny T, Abarbanel HDI. Spike-timing-dependent plasticity of inhibitory synapses in the entorhinal cortex. J Neurophysiol.
2006;96(6):3305–13.
11. D’Amour JA, Froemke RC. Inhibitory and excitatory spike-timing-dependent plasticity in the auditory cortex. Neuron. 2015;86(2):514–28.
12. Lazar A, Pipa G, Triesch J. SORN: a self-organizing recurrent neural network. Front Comp Neurosci. 2009;3:23.
Page 35 of 112
4.
P45
Simulation of EMG recording from hand muscle due to TMS
of motor cortex
Bahar Moezzi1, Nicolangelo Iannella1,4, Natalie Schaworonkow2, Lukas
Plogmacher2, Mitchell R. Goldsworthy3, Brenton Hordacre3, Mark D.
McDonnell1, Michael C. Ridding3, Jochen Triesch2
1
Computational and Theoretical Neuroscience Laboratory, School
of Information Technology and Mathematical Sciences, University
of South Australia, Australia; 2Frankfurt Institute for Advanced Studies,
Goethe‑Universität, Germany; 3Robinson Research Institute, School
of Medicine, University of Adelaide, Australia; 4School of Mathematical
Sciences, University of Nottingham, UK
Correspondence: Bahar Moezzi ‑ bahar.moezzi@unisa.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P45
Single pulse transcranial magnetic stimulation (TMS) is a technique
which (at moderate intensities) activates corticomotor neuronal output cells transynaptically and evokes a complex descending volley in
the corticospinal tract. Rusu et al. developed a computational model
of TMS induced I-waves that reproduced observed epidural recordings in conscious humans [1]. In humans, epidural responses can be
recorded in anaesthetized subjects during surgery or conscious subjects with electrodes implanted for the treatment of chronic pain. Such
opportunities are uncommon and invasive. The effects of TMS can be
non-invasively studied using surface electromyography (EMG) recordings from the hand first dorsal interosseous (FDI) muscle.
We simulated the surface EMG signal due to TMS of motor cortex in
the hand FDI muscle. Our model comprises a population of cortical
layer 2/3 cells, which drive layer 5 cortico-motoneuronal cells with
excitatory and inhibitory synaptic inputs as in [1]. The layer 5 cells in
turn project to a pool of motoneurons, which are modeled as an inhomogeneous population of integrate-and-fire neurons to simulate
motor unit recruitment and rate coding. The input to motoneurons
from cortical layer 5 consists of TMS-induced spikes and baseline firing. We modeled baseline firing with a Poisson drive to layer 2/3 cells.
Hermite-Rodriguez functions were used to simulate motor unit action
potential shape. The EMG signal was obtained from the summation of
motor unit action potentials of active motor units. Parameters were
tuned to simulate recordings from the FDI muscle.
Our simulated EMG signals match experimental surface EMG recordings due to TMS of motor cortex in the hand FDI muscle in shape,
size and time scale both at rest and during voluntary contraction (see
Fig. 31 Comparison of simulated and experimental EMG during A
rest, B 10 % maximum voluntary contraction
Fig. 31). The simulated EMG traces exhibit cortical silent periods (CSP)
that lie within the biological range.
Reference
1. Rusu CV, Murakami M, Ziemann U, Triesch J. A model of TMS-induced
I-waves in motor cortex. Brain Stimul. 2014;7(3):401–14.
P46
Structure and dynamics of axon network formed in primary cell
culture
Martin Zapotocky1,2, Daniel Smit1,2,3, Coralie Fouquet3, Alain Trembleau3
1
Institute of Physiology of the Czech Academy of Sciences, Prague,
Czech Republic; 2Institute of Biophysics and Informatics, First Faculty
of Medicine, Charles University in Prague, Czech Republic; 3IBPS,
Neuroscience Paris Seine, CNRS UMR8246, Inserm U1130, UPMC UM 119,
Université Pierre et Marie Curie, Paris, France
Correspondence: Martin Zapotocky ‑ zapotocky@biomed.cas.cz
BMC Neuroscience 2016, 17(Suppl 1):P46
Axons growing in vivo or in culture may adhere to each other and form
a connected network, which subsequently guides the paths of newly
arriving axons. We investigated the development of such a network
formed by growing axons in primary cell culture.
Olfactory epithelium explants from mouse embryos (day 13–14) were
cultured on laminin substrate for 2 days and then recorded using DIC
or phase contrast videomicroscopy for up to 24 h. The growing axons
established a dense network within which large fascicles of axons
were progressively formed. Within the recorded time period, the network remained stable, with limited further gowth of the axons but
with ongoing rearrangement in the network structure. Based on segmentation of the recorded images, we determined the principal network characteristics (including the total length, the total number of
vertices, and the network anisotropy) and their evolution in time.
This quantitative characterization permitted an analysis of the mechanisms of the observed network coarsening. We relate the network
dynamics to the elementary processes of zippering, during which
two axons or axon fascicles progressively adhere to each other [1]. We
compare the structural features of the network (such as the distribution of vertex angles) with those reported in an electron microscopy
investigation of a plexus of sensory neurites in Xenopus embryo [2].
We show that both our ex vivo study and the in vivo study of Ref. [2]
support a similar underlying mechanism of the formation of the axon
network.
Acknowledgements: Work supported by GAČR 14-16755S, GAUK
396213, MŠMT 7AMB12FR002, NIH 1RO1DCO12441 and ANR
2010-BLAN-1401-01.
References
1. Smit D, Fouquet C, Pincet F, Trembleau A, Zapotocky M. Axon zippering
in neuronal cell culture and its biophysical modeling. BMC Neurosci.
2015;16(Suppl. 1):P298.
BMC Neurosci 2016, 17(Suppl 1):54
2.
Roberts A, Taylor JSH. A scanning electron microscope study of the development of a peripheral sensory neurite network. J Embryol Exp Morph.
1982;69:237–50.
P47
Efficient signal processing and sampling in random networks
that generate variability
Sakyasingha Dasgupta1,2, Isao Nishikawa3, Kazuyuki Aihara3, Taro
Toyoizumi2
1
IBM Research ‑ Tokyo, Tokyo, Japan; 2RIKEN Brain Science Institute, Tokyo,
Japan; 3The University of Tokyo, Tokyo, Japan
Correspondence: Sakyasingha Dasgupta ‑ sdasgup@jp.ibm.com
BMC Neuroscience 2016, 17(Suppl 1):P47
The source of cortical variability and its influence on signal processing
remain an open question. We address the latter, by studying two types
of randomly connected networks of quadratic integrate-and-fire neurons
with balanced excitation-inhibition that produce irregular spontaneous
activity patterns (Fig. 32A): (a) a deterministic network with strong synaptic interactions that actively generates variability by chaotic dynamics
(internal noise) and (b) a stochastic network that has weak synaptic interactions but receives noisy input (external noise), e.g. by stochastic vesicle
releases. These networks of spiking neurons are analytically tractable in
the limit of a large network-size and slow synaptic-time-constant. Despite
the difference in their sources of variability, spontaneous (baseline) activity patterns of these two models are indistinguishable unless majority
of neurons are simultaneously recorded. We characterize the network
behavior with dynamic mean field analysis and reveal a single-parameter
family that allows interpolation between the two networks, sharing nearly
identical spontaneous activity (Fig. 32B). Despite the close similarity in the
spontaneous activity, the two networks exhibit remarkably different sensitivity to external stimuli. Input to the former network reverberates internally and can be successfully read out over long time. Contrarily, input to
the latter network rapidly decays and can be read out only for short time.
This is also observed in the significant changes in the spiking probability
of evoked responses across this family (Fig. 32C). The difference between
Fig. 32 A Schematic illustrations of the two balanced QIF networks
models considered in the present study. The left network consists
of strongly coupled neurons without noise, while the right network
consists of weak coupling among neurons with noisy input. B Nearly
identical rate autocorrelation functions in the two networks. The red
line (C0) represents the value of the autocorrelation at time 0 and cyan
line (C∞) is the value of auto-correlation function in the limit of large
t. C Change in spiking probability for different network connectivity
strengths (g̃), after being stimulated by a brief input at time t = 0
Page 36 of 112
the two networks is further enhanced if input synapses undergo activitydependent plasticity, producing significant difference in the ability to
decode external input from neural activity. We show that, this difference
naturally leads to distinct performance of the two networks to integrate
spatio-temporally distinct signals from multiple sources. Unlike its stochastic counterpart, the deterministic chaotic network activity can serve
as a reservoir to perform near optimal Bayesian integration and MonteCarlo sampling from the posterior distribution. We describe implications
of the differences between deterministic and stochastic neural computation on population coding and neural plasticity.
P48
Modeling the effect of riluzole on bursting in respiratory neural
networks
Daniel T. Robb1, Nick Mellen2, and Natalia Toporikova3
1
Department of Mathematics, Computer Science and Physics, Roanoke
College, Salem, VA 24153, USA; 2Department of Pediatrics, University
of Louisville, Louisville, KY 40208, USA; 3Department of Biology,
Washington and Lee University, Lexington, VA 24450, USA
Correspondence: Daniel T. Robb ‑ robb@roanoke.edu
BMC Neuroscience 2016, 17(Suppl 1):P48
To accommodate constantly changing environmental and metabolic
demands, breathing should be able to vary flexibly within a range of frequencies. The respiratory neural network in the pre-Botzinger complex
of the ventrolateral medulla controls and flexibly maintains the breathing rhythm, coordinating network-wide bursting to signal the inspiratory
phase of the breath. The frequency of this rhythmic activity is controlled
by a number of neuromodulators, the majority of which are excitatory.
Therefore, the central pattern generator for rhythmic respiratory activity
should possess two seemingly contradictory properties: it has to be able
to change frequency in response to excitatory input, but it also has to
preserve stable rhythmic activity under a wide range of conditions.
A persistent sodium current (INaP) been identified as one of the key currents for generation of inspiratory activity [1]. It has been shown that
some of the neurons in Pre-BotC possess an intrinsic bursting mechanism, which relies on inactivation of this current. Higher expression
of INaP correlates with higher burst frequency of a single pacemaker
neuron [2]. However, the INaP pacemaker mechanism can only function within very narrow ranges of external excitation—NaP dependent pacemaker tends to switch to tonic firing after a small increase in
depolarizing current [3].
In this combined experimental and computational study, we tested
the effect of the persistent sodium blocker Riluzole (RIL) in several different levels of continuous depolarization, simulated by application of
K+. Whereas increased potassium increases the bursting frequency of
the control network, in the presence of RIL the increased potassium
does not alter the bursting frequency (Fig. 33). These findings indicate
Fig. 33 Summary of experiment on the effect of riluzole on the
dependence of burst frequency on potassium concentration.
Without riluzole (left), the frequency increases steadily with increasing
potassium concentration. With riluzole present (right), the frequency
remains essentially constant with increasing potassium concentration
BMC Neurosci 2016, 17(Suppl 1):54
that INaP is responsible for flexible modulation of respiratory rhythm,
but there is another mechanism, which can sustain rhythmic activity in
its absence. We developed a computational model which incorporates
a Calcium sensitive Non-specific cationic current (IcaN) in addition to
INaP. Our simulations indicate that IcaN and INaP can maintain the rhythm
in respiratory neurons in the presence of RIL, and are capable of providing stable oscillations in the presence of tonic excitation by K+.
References
1. Butera RJ Jr, Rinzel J, Smith JC. Models of respiratory rhythm generation in
the pre-Bötzinger complex. I. Bursting pacemaker neurons. J Neurophysiol. 1999;82:382–97.
2. Purvis LK, Smith JC, Koizumi H, Butera RJ. Intrinsic bursters increase the
robustness of rhythm generation in an excitatory network. J Neurophysiol. 2007;97:1515–26.
3. Del Negro CA, Morgado-Valle C, Hayes JA, Mackay DD, Pace RW, Crowder
EA, Feldman JL. Sodium and calcium current-mediated pacemaker neurons and respiratory rhythm generation. J Neurosci Off J Soc Neurosci.
2005;25:446–53.
P49
Mapping relaxation training using effective connectivity analysis
Rongxiang Tang1, Yi‑Yuan Tang2
1
Department of Psychology, Washington University in St. Louis, St. Louis,
MO 63130, USA; 2Department of Psychological Sciences, Texas Tech
University, TX 79409, USA
Correspondence: Yi‑Yuan Tang ‑ yiyuan.tang@ttu.edu
BMC Neuroscience 2016, 17(Suppl 1):P49
Relaxation training (RT)is a behavioral therapy that has been applied
in stress management, muscle relaxation and other health benefit.
However, compared to short-term meditation training, previous studies did not show the significant differences in brain changes following
same amount of RT [1,2]. One possible reason might derive from the
insensitive correlation based routine functional connectivity method
that could not reveal training-related changes in effective connectivity (directed information flow) among these distributed brain regions.
Here, we applied a novel spectral dynamic causal modeling (spDCM)
to resting state fMRI to characterize changes in effective connectivity.
Twenty-three healthy college students were recruited through campus
advertisements and received 4 weeks of RT (10 h in total), previously
reported in our randomized studies [1, 2]. All neuroimaging data were collected using an Allegra 3-Telsa Siemens scanner and processed using the
Data Processing Assistant for Resting-State fMRI, which is based on SPM
and Resting-State fMRI Data Analysis Toolkit [3]. For each participant, the
subsequent standard procedures included slice timing, motion correction, regression of WM/CSF signals, and spatial normalization [3]. Based
on previous literature, we specified four regions of interest within default
mode network (DMN)—medial prefrontal cortex (mPFC), posterior cingulate cortex (PCC), and bilateral inferior parietal lobule (left IPL and right
IPL), same coordinates as in previous spDCM studies [4]. A standard DCM
analysis involves a specification of plausible models, which are then allows
the model parameters to be estimated following Bayesian model selection. In both pre- and post-RT conditions, the procedure selected the fully
connected model as the best model with a posterior probability of almost
1. The fully connected model had 24 parameters describing the extrinsic connections between nodes, the intrinsic (self-connections) within
nodes and neuronal parameters describing the neuronal fluctuations
within each node. We used Bayesian Parametric Average to quantify the
differences between pre- and post-RT, and a classical multivariate test—
canonical variate analysis to test for any significances in these differences
[4]. Our results showed no significant differences in causal relationships
among the above nodes following RT (all P > 0.05).
Conclusions Four weeks of RT could not induce significant changes
in effective connectivity among DMN nodes. Long-term RT effect on
brain changes warrants further investigation.
Acknowledgements: This work was supported by the Office of Naval
Research.
Page 37 of 112
References
1. Tang YY, Holzel BK, Posner MI. The neuroscience of mindfulness meditation. Nat Rev Neurosci. 2015;16:213–25.
2. Tang YY, Lu Q, Geng X, Stein EA, Yang Y, Posner MI. Short-term meditation
induces white matter changes in the anterior cingulate. Proc Natl Acad
Sci USA. 2010;107:15649–52.
3. Tang YY, Tang R, Posner MI. Brief meditation training induces smoking
reduction. Proc Natl Acad Sci USA. 2013;110:13971–75.
4. Razi A, Kahan J, Rees G, Friston KJ. Construct validation of a DCM for resting state fMRI. Neuroimage. 2015;106:1–14.
P50
Modeling neuron oscillation of implicit sequence learning
Guangsheng Liang1, Seth A. Kiser2,3, James H. Howard, Jr.3, Yi‑Yuan Tang1
1
Department of Psychological Sciences, Texas Tech University, TX 79409,
USA; 2The Department of Veteran Affairs, District of Columbia VA Medical
Center, Washington, DC 20420, USA; 3Department of Psychology, The
Catholic University of America, Washington, DC 20064, USA
Correspondence: Yi‑Yuan Tang ‑ yiyuan.tang@ttu.edu
BMC Neuroscience 2016, 17(Suppl 1):P50
Implicit learning (IL) occurs without goal-directed intent or conscious
awareness but has important influences on our everyday functioning and overall health such as environmental adaptation, developing
habits and aversions. Most of IL studies used event-related potentials (ERPs) to study brain response by taking the grand average of all
event-related brain signals. How neuron oscillation (EEG frequency
band) involves in IL remains unknown. Moreover, ERP analysis requires
brain signals that are not only time locked, but also phase locked to
the event, therefore the information with phase locked signals are
missed and not presented in potentials. To address this issue, we
applied time–frequency analysis and cluster-based permutation test
in this study.
Fifteen healthy participants were recruited to perform three sessions of triplets learning task (TLT), an IL task commonly used in the
field [1]. Three successive cues were presented and participants were
asked to observe the first two cues and only respond to the third cue
(target) by pressing corresponding keys. During the task, EEG signals
were recorded. Cluster based permutation on alpha and theta band
is used to deal with family-wise error rate and in the same time, help
to find out difference occurred in specific time range along with spatial information among different triplet types.
Base on the behavioral result, overall learning occurs in session1,
while triplet-specific learning takes place in session2. We find significant difference in both alpha (8–13 Hz) and theta (4–8 Hz) frequency
band. For alpha band, power modulation shows significant difference between high versus low frequency triplet group in session2
in the frontal cortex. For theta band, theta power shows significant
difference between session1 and session3 in the frontal cortex. It
started from as early as target onset until the end of the trial in high
frequency triplet group. However, in the low frequency triplet group,
the power differential occurs later, from around 1000 ms till the end
of the next trial.
Conclusions Behavioral result showed that the brain learned the regularity of sequence implicitly. Alpha power modulation indicated that
the brain allocated resource in attention among two different triplet
types. Theta power modulation showed the difference of memory
processing and retrieval among two different triplet types. Our results
indicated that participants did not find the regularity of the triplet
types till the end of the study, but the brain in fact reacts to these two
different triplet types differently.
Acknowledgements: This work was supported by the Office of Naval
Research.
Reference
1. Howard JH, Howard DV, Dennis N, Kelly AJ. Implicit learning of predictive
relationships in three-element visual sequences by young and old adults.
J Exp Psychol Learn Mem Cogn. 2008, 34: 1139–57.
BMC Neurosci 2016, 17(Suppl 1):54
P51
The role of cerebellar short‑term synaptic plasticity in the
pathology and medication of downbeat nystagmus
Julia Goncharenko1, Neil Davey1, Maria Schilstra1, Volker Steuber1
1
Centre for Computer Science and Informatics Research, University
of Hertfordshire, Hatfield, AL10 9EJ, UK
Correspondence: Julia Goncharenko ‑ i.goncharenko@herts.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P51
Downbeat nystagmus (DBN) is a common eye fixation disorder that is
linked to cerebellar pathology. DBN patients are treated with 4-aminopyridine (4-AP), a K channel blocker, but the underlying mechanism is
unclear. DBN is associated with an increased activity of floccular target neurons (FTNs) in the vestibular nuclei. It was previously believed
that the reason for the increased activity of FTNs in DBN is a pathological decrease in the spike rate of their inhibitory Purkinje cell inputs,
and that the effect of 4-AP in treating DBN could be mediated by an
increased Purkinje cell activity, which would restore the inhibition
of FTNs and bring their activity back to normal [1]. This assumption,
however, has been questioned by in vitro recordings of Purkinje cells
from tottering (tg/tg) mice, a mouse model of DBN. It was shown that
therapeutic concentrations of 4-AP did not increase the spike rate of
the Purkinje cells, but that they restored the regularity of their spiking,
which is impaired in tg/tg mice [2].
Prompted by these experiments, Glasauer and colleagues performed
computer simulations to investigate the effect of the regularity of
Purkinje cell spiking on the activity of FTNs [3]. Using a conductance based FTN model, they found that changes in the regularity of
the Purkinje cell input only affected the FTN spike rate when the
input was synchronized. In this case, increasing the regularity of the
Purkinje cell spiking resulted in larger gaps in the inhibitory input to
the FTN and an increased FTN spike rate. These results predict that the
increased irregularity in the Purkinje cell activity in DBN should lead to
a decreased activity of the FTNs, rather than the increased activity that
is found in experiments, and they are therefore unable to explain the
therapeutic effect of 4-AP.
However, the model by Glasauer and colleagues does not take
short-term depression (STD) at the Purkinje cell—FTN synapses into
account. We hypothesized that this absence of STD could explain
the apparent contradiction between the experimental [2] and computational [3] results. To study the role of STD in the pathology and
4-AP treatment of DBN, we used a morphologically realistic conductance based model of a cerebellar nucleus (CN) neuron [4, 5] as an
FTN model to simulate the effect of irregular versus regular Purkinje
cell input. The coefficients of variation of the irregular and regular Purkinje cell spike trains during DBN and after 4-AP treatment,
respectively, were taken from recordings from wild-type and tg/tg
mice [6], which served as a model system for DBN. We presented the
FTN model with synchronized and unsynchronized input and found
that, for both conditions, irregular (DBN) input trains resulted in
higher FTN spike rates than regular (4-AP) ones. In the presence of
unsynchronized Purkinje cell input, the acceleration of the FTN spike
output during simulated DBN and the deceleration during simulated
4-AP treatment depended on STD at the Purkinje cell synapses. Our
results provide a potential explanation for the pathology and 4-AP
treatment of pathological nystagmus.
References
1. Glasauer S, Kalla R, Buttner U, Strupp M, Brandt T. 4-aminopyridine
restores visual ocular motor function in upbeat nystagmus. J Neurol
Neurosurg Psychiatry. 2005;76:451–3.
2. Alvina K, Khodakhah K. The therapeutic mode of action of 4-aminopyridine in cerebellar ataxia. J Neurosci. 2010;30:7258–68.
3. Glasauer S, Rössert C, Strupp M. The role of regularity and synchrony of
cerebellar Purkinje cells for pathological nystagmus. Ann NY Acad Sci.
2011;1233:162–7.
4. Steuber V, Schultheiss NV, Silver RA, de Schutter E, Jaeger D. Determinants of synaptic integration and heterogeneity in rebound firing
explored with data-driven models of deep cerebellar nucleus cells. J
Comp Neurosci. 2011;30:633–58.
Page 38 of 112
5.
6.
Luthman J, Hoebeek FE, Maex R, Davey N, Adams R, de Zeeuw CI, Steuber V. STD-dependent and independent encoding of input irregularity
as spike rate in a computational model of a cerebellar nucleus neuron.
Cerebellum. 2011;10:667–82.
Hoebeek FE, Stahl JS, van Alphen AM, Schonewille M, Luo C, Rutteman M, van den Maagdenberg AM, Molenaar PC, Goossens HH, Frens
MA, et al. Increased noise level of Purkinje cell activities minimizes
impact of their modulation during sensorimotor control. Neuron 2005,
45(6):953–965.
P52
Nonlinear response of noisy neurons
Sergej O. Voronenko1,2, Benjamin Lindner1,2
1
Department of Physics, Humboldt University, Berlin, 10099, Germany;
2
Bernstein Center for Computational Neuroscience, Berlin, 10115,
Germany
Correspondence: Sergej O. Voronenko ‑ sergej@physik.hu‑berlin.de
BMC Neuroscience 2016, 17(Suppl 1):P52
In many neuronal systems that exhibit high trial-to-trial variability
the time-dependent firing rate is thought to be the main information
channel for time-dependent signals. However, for nerve cells with low
intrinsic noise and highly oscillatory activity synchronization, mode
locking and frequency locking seem to be of major importance. Here,
we present an extension to the linear response theory [1, 2] for the
leaky integrate-and-fire neuron model to second order and demonstrate how the time-dependent firing rate can exhibit features that
are reminiscent of mode-locking and frequency-locking. Although
our theory allows to predict the response to general weak timedependent signals, the second-order effects are best demonstrated
using cosine signals as in Fig. 34A. We consider a leaky integrate-andfire model for which the subthreshold voltage, Fig. 34B, is subject to
the signal and to Gaussian white noise. Whenever the voltage hits the
threshold, it is reset to zero and a spike time is recorded in the raster
plot, Fig. 34C. The firing rate can be obtained numerically by averaging over the spike trains or via a perturbation approach similar to the
weakly nonlinear analysis in [3]. We find that the firing rate can exhibit
pronounced nonlinear behavior as can be seen from the excitation of
a harmonic oscillation in Fig. 34D. Further effects that are not shown in
Fig. 34 but are revealed by our analysis are a signal-dependent change
of the mean firing rate and a pronounced nonlinear response to the
sum of two cosine signals.
Summary and conclusions Here we demonstrate that the timedependent firing rate (equivalent to the instantaneous population rate
for neurons driven by a common stimulus) can exhibit pronounced
Fig. 34 Nonlinear modulation of the firing rate by a cosine signal. A
Signal, B subthreshold voltage, C rasterplot, D The time-dependent
firing rate (red, noisy trace) is significantly different from the linear
theory (dashed line) but is accurately described by the second-order
response (solid line)
BMC Neurosci 2016, 17(Suppl 1):54
nonlinearities even for weak signal amplitudes. The linear theory does
not only give quantitatively wrong predictions but also fails to capture
the timing of the modulation peaks. Hence, our theory has not only
implications for sinusoidal stimulation that is commonly used to study
dynamic properties of nerve cells but also demonstrates the relevance
of the nonlinear response for the encoding of complex time-dependent signals.
Acknowledgements: This work was supported by the BMBF (FKZ:
01GQ1001A) and the DFG (research training group GRK1589/2).
References
1. Brunel N, Chance FS, Fourcaud N, Abbott LF. Effects of synaptic noise
and filtering on the frequency response of spiking neurons. PRL.
2001;86(10):2186–9.
2. Lindner B, Schimansky-Geier L. Transmission of noise coded versus additive signals through a neuronal ensemble. PRL. 2001;86(14):2934–7.
3. Brunel N, Hakim V. Fast global oscillations in networks of integrate-andfire neurons with low firing rates. Neural Comput. 1999;11(7):1621–71.
P53
Behavioral embedding suggests multiple chaotic dimensions
underlie C. elegans locomotion
Tosif Ahamed1, Greg Stephens1,2
1
Biological Physics Theory Unit, Okinawa Institute of Science
and Technology, Okinawa 904‑0495, Japan; 2Department of Physics
and Astronomy, Vrije Universiteit Amsterdam
Correspondence: Tosif Ahamed ‑ tosif.ahamed@oist.jp
BMC Neuroscience 2016, 17(Suppl 1):P53
Behavior is the primary output of an organism; genetic and neural circuits, no matter how complex, seek to optimize this output. A
Page 39 of 112
quantitative understanding of behavior is therefore crucial to our
understanding of biological processes. A key characteristic of natural
behavior is variability; even the most stereotyped movements such
as reaching to a target, which are similar in aggregate, can vary substantially from trial to trial. In motor control such variability is often
ascribed to noise in the sensorimotor control circuit. On the other
hand, deterministic dynamical systems can generate variability intrinsically when operating in a chaotic regime. Differentiating between
the two is important as they generate separate mechanistic predictions about how variability is generated in the brain. Here, we use
tools from nonlinear dynamics to understand behavioral variability
in the movement of C. elegans. We reconstruct a 6-dimensional phase
space by developing a novel extension of multivariate singular systems analysis [1] and applying it to a low-dimensional but complete
representation of worm postures obtained from videos of freely foraging worms [2]. At a coarse level, the reconstructed phase space naturally separates into three stereotyped behaviors: forward locomotion,
reversals and turns (Fig. 35A, B). However, there is also substantial
variability at finer scales, which is reflected in positive maximal Lyapunov exponents (MLE) [3] within trajectories corresponding to each
individual behavior (Fig. 35C). The MLEs calculated this way differ significantly from MLEs calculated from a random shuffle of the data that
preserves its linear structure (or power spectrum). This implies that the
positive MLEs, which indicate sensitive dependence to initial conditions in C. elegans behavior are a result of nonlinear structure present
in the dynamics. These results are strengthened by the fact that we
observe little inter-animal variability in the estimated values of MLE,
additionally the values also agree well with the estimated time scales
of the three behaviors. Based on these observations we propose that
C. elegans behavior might be driven by the activity of multiple coupled chaotic attractors. We expect our analysis will also be relevant in
understanding global neural dynamics, recently imaged in freely-moving worms [4].
References
1. Read PL. Phase portrait reconstruction using multivariate singular systems analysis. Phys D. 1993;69(3):353–65.
2. Stephens GJ, Johnson-Kerner, B, Bialek W, Ryu WS. Dimensionality and dynamics in the behavior of C. elegans. PLoS Comput Biol.
2008;4(4):e1000028.
3. Kantz H. A robust method to estimate the maximal Lyapunov exponent
of a time series. Phys Lett A. 1994;185(1):77–87.
4. Nguyen JP, Shipley FB, Linder AN, Plummer GS, Liu M, Setru SU, Shaevitz
JW, Leifer AM. Whole-brain calcium imaging with cellular resolution in
freely behaving Caenorhabditis elegans. PNAS. 2015;201507110.
Fig. 35 Phase space portrait and divergence of nearby trajectories.
A The top panel shows the orthogonal relationship between the
forward and reversal behaviors, while the bottom panel shows the
transition from reversal to an omega turn in the phase space. To aid
visualization color coding is done by radial distance from the origin.
B Escape response visualized in the phase planes. When the worm
is hit with a laser impulse, it makes a reversal, followed by an omega
turn and then resumes forward crawling. Color map encodes time
in frames. C Divergence curves for the three different attractors.
Y-axis shows the exponential of the divergence between neighboring trajectories plotted on a semilog scale on the y axis, each curve
corresponds to a single worm (n = 12). λL is estimated by calculating
the slope of the linear region. Boxplots show the range of λL obtained
from different animals
P54
Fast and scalable spike sorting for large and dense
multi‑electrodes recordings
Pierre Yger1, Baptiste Lefebvre1, Giulia Lia Beatrice Spampinato1, Elric
Esposito, Marcel Stimberg et Olivier Marre1
1
Institut de la Vision, INSERM UMRS 968, CNRS UMR 7210, Paris
Correspondence: Pierre Yger ‑ pierre.yger@inserm.com
BMC Neuroscience 2016, 17(Suppl 1):P54
Understanding how assemblies of neurons encode information
requires recording of large populations of cells in the brain. In recent
years, multi-electrode arrays and large silicon probes have been developed to record simultaneously from thousands of electrodes packed
with a high density. However, these new devices challenge the classical way to do spike sorting. First, the large number of electrodes
preclude approaches based on manual clustering. Even automatic
approaches need to be fast enough to handle the amount of extracellular data. Second, the density of the electrodes is high enough so that
a single spike will be detected on many electrodes. So the different
channels must be processed simultaneously. Third, within a large and
dense array of electrodes, overlapping spikes are rather the rule than
the exception, and it is known that classical clustering methods cannot easily capture the synchronous occurrence of two spikes from two
different cells [1].
BMC Neurosci 2016, 17(Suppl 1):54
Here we developed a new software to solve all these aforementioned
issues, based on a highly automated algorithm to extract spikes from
extracellular data, and show that this algorithm reached near optimal
performance both in vitro and in vivo. The algorithm is composed of
two main steps: (1) a “template-finding” phase to extract the cell templates, i.e. the pattern of activity evoked over many electrodes when
one neuron fires an action potential; (2) a “template-matching” phase
where the templates are matched onto the raw data to find the location of the spikes. The manual intervention by the user is reduced to
the minimal, and the time spent on manual curation did not scale with
the number of electrodes. For the template-finding phase, we start by
detecting all the possible times in the raw data that could contain a
spike. Spikes are then clustered into groups using a density-based
clustering derived from [1], and we then extract the template corresponding to each group. In the fitting phase, we match the templates
onto the raw data with a method that allows amplitude variation for
each template [2]. The algorithm is written in Python and is entirely
parallelized such that it can handle large amount of data. It also provides a graphical user interface so that the output of the algorithm can
be checked, and to refine the sorting.
We tested our algorithm with large-scale data from in vitro and in vivo
recordings, from 32 and up to 4225 electrodes. In all cases, we estimated its performance on data with ground truth, i.e. cases where the
solution to the sorting problem is at least partially known. The performance was always close to the maximal expected performance. Therefore, our method appears as a general solution to sort spikes from
large-scale extracellular recordings.
References
1. Einevoll GT, et al. Towards reliable spike-train recordings from thousands
of neurons with multielectrodes. Curr Opin Neurobiol. 2012;22:11–17.
2. Rodriguez A, et al. Clustering by fast search and find of density peaks.
Science. 2014;344(6191):1492–96.
P55
Sufficient sampling rates for fast hand motion tracking
Hansol Choi1, Min‑Ho Song2
1
Bernstein Center Freiburg, Institute of Biology III, University of Freiburg,
Germany, 79100; 2fourMs group, Dept. Musicology, University of Oslo,
Norway, 0371
Correspondence: Min‑Ho Song ‑ minho.song@imv.uio.no
BMC Neuroscience 2016, 17(Suppl 1):P55
When tracking fine motor behaviors in human body parts, passive
marker-based tracking is one of the best-suited methods not only
because of its high spatial precision and temporal resolution, but also
allowing high degrees-of-freedom [1]. However, the passive marker
approach suffers from identity confusion problem (Fig. 36A) between
Fig. 36 Experimental design and result. A Marker confusion. Grey
dots are markers. d1, d2 are the distances between markers, ts is the
sampling latency, v is speed of marker. Green lines show the markers,
which identified as same. Left correct identification right example of
marker confusion. B Experimental set up. Red dots are keys to press by
the thumb and the little finger during repeats. C The probabilities of
continuous marker identification
Page 40 of 112
the markers. As the speed of motion increases, sufficient sampling rate
is required to avoid the problem. In a recent study [2], we reported
that the problem still occurs even with the sampling rate significantly
higher than the Nyquist sampling rate. The study suggested a sampling
rate criterion to avoid identity problem for the worst-case condition.
In this poster, the confusion problem is tested in more realistic human
motor control behavior. Grids of 3 × 3 markers with different distances
(1, 1.5 and 2 cm) were attached to a skilled piano player’s right hand
(Fig. 36B). The experimental task was repeated right-hand alternative keystrokes between D#5 and D#7 (two octave) with a tempo of
176 bpm for 10 s. This is an excerpt from Liszt’s La Campanella, which
requires fast horizontal jump of the right-hand. These motions were
recorded with 7 optical motion capture cameras (Qualisys Ltd. Oqus
400) changing the sampling rates from 50 to 200 Hz. The maximum frequency components of these hand movements were lower than 8 Hz.
The probability of successful tracking is measured by counting the
number of successful repetition of the center marker (Fig. 36C). Estimated required sampling rates for successful tracking (where the
probabilities reach 100 %) were 101, 137z, and 181 Hz (fitted to piecewise linear functions by expectation maximization). The theoretically
predicted values are 176, 235, and 353 Hz [2].
We found that the required sampling rates are lower than the theoretical criterion. This is because the theoretical prediction was developed
to avoid the worst case where marker trajectories overlap from perfect
periodic motion; not realistic for human movement, which has variability. Our results show that in practical situations involving human
movements, the sampling criterion can be weakened considerably.
But, it should be note that a motion slower than 10 Hz still requires
more than 100 Hz, which far exceeds the Nyquist sampling rate.
References
1. Guerra-Filho G. Optical motion capture: theory and implementation. J
Theor Appl Inf. 2005;12(2):61–8.9
2. Song M-H, Godøy RI. How fast is your body motion? Determining a
sufficient frame rate for an optical motion tracking system using passive
markers. PLoS One (in press).
P56
Linear readout of object manifolds
SueYeon Chung1, Dan D. Lee2, Haim Sompolinsky1,3
1
Center for Brain Science, Harvard University, Cambridge, MA 02138,
USA; 2Department of Electrical and Systems Engineering, University
of Pennsylvania, Philadelphia, PA 19104, USA; 3Edmond and Lily Safra
Center for Brain Sciences, Hebrew University, Jerusalem 91904, Israel
Correspondence: SueYeon Chung ‑ schung@fas.harvard.edu
BMC Neuroscience 2016, 17(Suppl 1):P56
Objects are represented in sensory systems by continuous manifolds due to sensitivity of neuronal responses to changes in physical
features such as location, orientation, and intensity [1]. It has been
hypothesized that object identity can be decoded from high level representations, by simple downstream readout networks. What makes
certain sensory representations better suited for invariant decoding
of objects by downstream networks? We generalize Gardner’s statistical mechanical analysis of points [2, 3] and establish a replica theory of
linear classification of manifolds synthesizing statistical and geometric properties of high dimensional signals. We show how changes in
the dimensionality, size, and shape of the object manifolds affect the
capacity and the distribution of configurations in downstream perceptrons (Fig. 37).
Our analysis shows how linear separability of the manifolds depends
intimately upon the dimensionality, size and shape of the the manifolds. These properties are expected to differ at different stages in the
sensory hierarchy. Thus, the present work enables systematic analysis
of the degree to which this reformatting enhances the capacity for
object classification in different sensory processing stages. The present
work lays the groundwork for a computational theory of neuronal processing of objects in the presence of variability, providing quantitative
measures for assessing the properties of representations in biological
and artificial neural networks.
BMC Neurosci 2016, 17(Suppl 1):54
Page 41 of 112
and Archaerhodopsin, respectively, into our conductance based PBC
models. A library of model PBC neurons was generated by varying the
conductance of each ion channel over equally spaced intervals. PRCs
were then calculated for each neuron capable of producing rhythmic
bursting. Preliminary results found that in general depolarizing perturbations produced qualitatively similar PRCs and could both advance
or delay the next cycle. Conversely, hyperpolarizing perturbations produced qualitatively distinct PRCs depending on the combination of
conductance magnitudes. In conclusion, these PRCs provide a method
for differentiating models of intrinsic bursting and rhythm generation
based on underlying biophysical mechanisms and provide a means for
interpreting experimentally derived PRCs from the PBC.
Fig. 37 Theoretical predictions (lines) and numerical simulation
(markers) are shown. A1 Classification of line segments. (Solid) lines
embedded in the margin, (dotted) lines touching the margin, (striped)
interior lines .A2 Capacity α = P/N of a network N = 200 as a function
of R (line length) with margins κ = 0, 0.5. A3 Fraction of configurations at capacity with κ = 0. (red) lines in the margin, (blue) touching
the margin, (black) interior lines. B1 D2 balls, B2 capacity √
α = P/N
for κ = 0 for large D = 50 and R ∝√
D−1/2 as a function of R D. (Blue
solid) αD(0, R) compared with α0(R D) (red square). (Inset) capacity α
at κ = 0 for 0.35 ≤ R ≤ 20 and D = 20: (blue) theoretical α compared
with approximate form (1 + R−2)/D (red dashed). C1 2D L1 balls. C2
Fraction of configurations as a function of radius R at capacity with
κ = 0. (red) entire manifold embedded, (blue) touching margin at a
single vertex, (gray) touching with two corners (one side), (purple)
interior manifold
References
1. DiCarlo JJ, Cox DD. Untangling invariant object recognition. Trends Cogn
Sci. 2007;11(8):333–41.
2. Gardner E. Maximum storage capacity in neural networks. EPL (Europhys
Lett). 1987;4(4):481
3. Abbott LF, Kepler TB. Universality in the space of interactions for network
models. J Phys A Math Gen. 1989;22(12):2031.
P57
Differentiating models of intrinsic bursting and rhythm
generation of the respiratory pre‑Bötzinger complex using phase
response curves
Ryan S. Phillips1,2, Jeffrey Smith1
1
NINDS, NIH, Bethesda, MD 20892, USA; 2Department of Physics,
University of New Hampshire, Durham, NH, 03824, USA
Correspondence: Ryan S. Phillips ‑ Ryan.Phillips@nih.gov
BMC Neuroscience 2016, 17(Suppl 1):P57
The pre-Bötzinger complex (PBC) is an essential rhythmogenic brainstem nucleus located in the ventrolateral medulla. Rhythmic output
from the PBC is relayed through premotor and motor neurons to the
diaphragm and intercostal muscles to drive the active inspiratory
phase of respiration. The specific biophysical mechanisms responsible
for generating rhythmic bursting and network synchronization are not
well understood and remain a highly controversial topic within the
field. A wide variety of biophysical mechanisms have been proposed to
explain the origins of intrinsic bursting and rhythmogenesis including
persistent sodium currents [1, 2], calcium-activated nonspecific cation
channels [2, 3], inositol trisphosphate (IP3) signaling [2], and synaptic
mechanisms [4]. Computational simulations of these models produce
similar patterns of bursting and network synchronization compared
to each other and experimental recordings despite having different
underlying mechanisms. In this theoretical study we demonstrate a
method to differentiate between biophysically distinct models using
phase response curves (PRCs). PRCs characterize the change in phase
of an oscillator as a function of the timing of a perturbation to the system. Depolarizing and hyperpolarizing perturbations were generated
by incorporating the light sensitive channels, Channelrhodopsin-2
References
1. Butera RJ, Rinzel J, Smith JC. Models of respiratory rhythm generation in
the pre-Bötzinger complex. I. Bursting pacemaker neurons. J Neurophysiol. 1999;82:382–97.
2. Jasinski PE, Molkov YI, Shevtsova NA, Smith JC, Rybak IA. Sodium and
calcium mechanisms of rhythmic bursting in excitatory neural networks
of the pre-Bötzinger complex: a computational modelling study. Eur J
Neurosci. 2013;37:212–30.
3. Rubin JE, Hayes JA, Mendenhall JL, Del Negro CA. Calcium-activated
nonspecific cation current and synaptic depression promote networkdependent burst oscillations. Proc Natl Acad Sci USA. 2009;106:2939–44.
4. Guerrier C, Hayes JA, Fortin G, Holcman D. Robust network oscillations
during mammalian respiratory rhythm generation driven by synaptic
dynamics. Proc Natl Acad Sci. 2015;201421997.
P58
The effect of inhibitory cell network interactions during theta
rhythms on extracellular field potentials in CA1
hippocampus
Alexandra Pierri Chatzikalymniou1,2, Katie Ferguson1,3, Frances K.
Skinner1,2,4
1
Krembil Research Institute, University Health Network, Toronto, ON,
Canada; 2Department of Physiology, University of Toronto, Toronto ON,
Canada; 3Department of Neuroscience, Yale School of Medicine, New
Haven, CT, 06520, USA; 4Department of Medicine (Neurology), University
of Toronto, Toronto ON, Canada
Correspondence: Alexandra Pierri Chatzikalymniou ‑ alexandra.chatzikalymniou@mail.utoronto.ca
BMC Neuroscience 2016, 17(Suppl 1):P58
Oscillatory local field potentials (LFPs) are extracellularly recorded potentials with frequencies of up to ~500 Hz. They are associated with a number of physiological functions in health and disease and complement the
information obtained by analysis of spikes. Because multiple neuronal
processes contribute to the LFP, the signal is inherently ambiguous and
more difficult to interpret than spikes [1]. However, the biophysical origin
of LFPs is well understood in the framework of volume conductor theory
[4]. Using “LFPy” [3], a python package that implements this framework,
we construct a pyramidal cell model in CA1 hippocampus which generates extracellular potentials. Our pyramidal cell model receives inhibitory
synaptic input from four different types of CA1 interneuron populations.
These interneuron models are taken from a previous, experimentally
constrained inhibitory network model developed to understand spontaneous theta (4–12 Hz) rhythms as expressed in an intact hippocampus
preparation [2]. We investigate the contribution of the different inhibitory cell type interactions to the extracellular potential. In our current
model we placed a virtual electrode probe along the vertical axis of the
pyramidal cell to record its output in a layer dependent manner. We identified distinct regimes where specific interneuron cell type interactions
distinctively affect the polarity, amplitude and frequency of the LFP signal (Fig. 38). We also distinguish between regimes where synaptic connection strengths preserve the extracellular potential frequency versus
those that lead to lag or abolishment of the extracellular rhythm. In this
way, our model helps us understand the cellular contributions to extracellular field patterns that arise in experimental recordings as a function
of biologically relevant network states when the efficacy of inhibitory
connections dynamically varies.
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 38 Example of the spatial attenuation of the extracellular potential signal for a particular set of inhibitory connections. The temporal
traces at two electrode locations are represented with blue and green
dots accordingly. Average over the absolute maximum extracellular
potential amplitudes is shown in 2D space. According to the schematic the rate of the extracellular signal spatial attenuation generated
by the pyramidal cell is approximately 400 μ
Acknowledgements: Supported by NSERC Canada, U. of T. Fellowship
P.S.L., and the SciNet HPC Consortium.
References
1. Buzsáki G, Anastassiou CA, Koch C. The origin of extracellular fields and
currents—EEG, ECoG, LFP and spikes. Nat Rev Neurosci. 2012;13:407–20.
2. Ferguson KA, Huh CYL, Amilhon B, Williams S, Skinner FK. Network
models provide insight into how oriens-lacunosum-moleculare (OLM)
and bistratified cell (BSC) interactions influence local CA1 theta rhythms.
Front Syst Neurosci. 2015;9:110
3. Lindén H, Hagen E, Leski S, Norheim ES, Pettersen KH, Einevoll GT. LFPy:
a tool for biophysical simulation of extracellular potentials generated by
detailed model neurons. Front Neuroinform. 2014;7:41.
4. Rall W, Shepherd GM. Theoretical reconstruction of field potentials and
dendrodendritic synaptic interactions in olfactory bulb. J Neurophysiol.
1968;31:884–915.
P59
Expansion recoding through sparse sampling in the cerebellar
input layer speeds learning
N. Alex Cayco Gajic1, Claudia Clopath2, R. Angus Silver1
1
Department of Neuroscience, Physiology and Pharmacology, University
College London, London, UK; 2Department of Bioengineering, Imperial
College London, London, UK
Correspondence: N. Alex Cayco Gajic ‑ natasha.gajic@ucl.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P59
Feed-forward networks often have many more output neurons than
input neurons. This is thought to enable them to project neuronal
activity patterns into a higher dimensional space. When such expansion recoding is combined with sparse representations it provides
a powerful way to increase the separation between activity patterns
[1–4]. In the input layer of the cerebellar cortex, granule cells (GCs)
integrate sensorimotor input from the less numerous mossy fibre
afferents (MFs), with each GC sampling from only 2 to 7 local MFs.
Recent work has revealed that this sparse sampling of input by GCs
provides an optimal tradeoff between information transmission and
sparsification over a range of activity levels [4]. Moreover, theories
of cerebellar function have linked expansion recoding under sparse
regimes to pattern separation and associative learning [5, 6]. However,
the relationship between the feedforward excitatory synaptic connectivity and the learning performance is poorly understood.
To investigate how the number of MF inputs per GC affects the performance we simulated a model of an 80µ ball of the MF-GC feedforward
layer with either random or clustered MF activation. The connectivity
Page 42 of 112
profile of the network was constrained with recent anatomical data [4].
MF stimulation was modeled as random binary patterns with varying
levels of population activity and correlation, and GCs as high-threshold rectified linear units. We then measured the speed at which granule cell population activity can be used to classify random patterns via
backpropagation learning as the number of synaptic inputs was varied. We found that the largest speedup of learning in the GC activity
(compared to learning based on the MF inputs) occurred when each
GC received only a few synaptic inputs. We probed this result by analyzing the eigenvalues of the covariance matrix of the population-level
activity, finding that sparse sampling of MF inputs allows GCs to both
expand and decorrelate MF activity patterns. Interestingly, this feature
is robustly preserved even in the presence of clustered inputs. In summary, we find that sparse sampling combined with sparsification of
activity allows GCs to optimize both pattern expansion and pattern
decorrelation.
Acknowledgement: This research is funded by the Wellcome Trust.
References
1. Laurent G. Olfactory network dynamics and the coding of multidimensional signals. Nat Rev Neurosci. 2001;3:884–95.
2. Olshausen BA, Field DJ. Sparse coding of sensory inputs. Curr Opin Neurobiol. 2004;14:481–7.
3. Billings G, Piasini E, Lorincz A, Nusser Z, Silver A. Network structure within
the cerebellar input layer enables lossless sparse encoding. Neuron.
2014;83:960–74.
4. Babadi B, Sompolinksy H. Sparseness and expansion in sensory representations. Neuron. 2014;83:1–14.
5. Marr D. A theory of cerebellar cortex. J Physiol. 1969;202:437–70.
6. Tyrell T, Wilshaw D. Cerebellar cortex: its simulation and the relevance of
Marr’s theory. Philos Trans R Soc B. 1992;336:239–57.
P60
A set of curated cortical models at multiple scales on Open Source
Brain
Padraig Gleeson1, Boris Marin1, Sadra Sadeh1, Adrian Quintana1, Matteo
Cantarelli2, Salvador Dura‑Bernal3, William W. Lytton3, Andrew Davison4, R.
Angus Silver1
1
Department of Neuroscience, Physiology and Pharmacology, University
College London, London, UK; 2Metacell LLC, San Diego, CA, USA; 3State
University of New York Downstate Medical Center, Brooklyn, NY, USA;
4
Neuroinformatics group Unité de Neurosciences, Information et
Complexité, CNRS, Gif sur Yvette, France
Correspondence: Padraig Gleeson ‑ p.gleeson@ucl.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P60
Computational models of spiking cortical networks are implemented
using a variety of approaches from large scale models with simplified
point neurons and anatomically inspired connectivity, to networks on
smaller scales with morphologically and biophysically detailed neurons. In between these scales many published models have used intermediate representations of neurons (e.g. conductance based with one
compartment or abstract morphologies). These studies, and the associated modelling scripts, provide many potential starting points for
experimental and theoretical neuroscientists wishing to use biologically constrained cortical models in their investigations. In addition,
there are an increasing number of public neuroinformatics resources
which are providing structured experimental data on the electrophysiology, connectivity and morphology of cortical neurons. While these
modelling and experimental resources should lead to a proliferation
in well constrained cortical models there remain a number of practical and technical barriers to more widespread development and use of
such models among researchers.
The Open Source Brain (OSB) initiative (http://www.opensourcebrain.
org) is a resource for collaborative development of models in computational neuroscience. Sharing of models in standardised representations such as NeuroML 2 [1] and PyNN is encouraged and actively
supported on the site. Conversion of cell and network models to
NeuroML allows them to be visualised and analysed in 3D in a standard web browser through the OSB website. We have recently added
BMC Neurosci 2016, 17(Suppl 1):54
a feature to allow simulations to be executed on our servers (e.g. by
conversion to NEURON) and the results displayed within the browser.
We have been actively converting published cortical models to NeuroML format and making these available on OSB. These range from
point neuron models [2, 3], to abstract [4] and detailed [5] multicompartmental models. We are working to develop these and others into a
curated set of cortical models in a common format which can be used
as the basis for new models. We have also developed frameworks for
importing resources from neuroinformatics datasets such as the Allen
Institute Cell Types database (http://celltypes.brain-map.org) and NeuroMorpho.org (http://neuromorpho.org). We have greatly improved
compatibility between PyNN and NeuroML, allowing the modeller
freedom to choose between procedural (Python) and declarative
(XML) model specification. We have also extended a model optimisation framework (https://github.com/NeuralEnsemble/NeuroTune)
facilitating generation of new NeuroML models from electrophysiological data. All of this work is aimed at making existing cortical models easier to access, visualise and simulate, simplifying development
of new models based on these prototypes, and ensuring the latest
experimental datasets can be used to constrain and validate complex
models of cortical function.
Acknowledgements: This work has been primarily funded by the
Wellcome Trust (101445/095667).
References
1. Cannon RC, Gleeson P, Crook S, Ganapathy G., Marin B, Piasini E, Silver RA.
LEMS: a language for expressing complex biological models in concise
and hierarchical form and its use in underpinning NeuroML2. Front
Neuroinform. 2014;8:79.
2. Izhikevich E. Simple model of spiking neurons. IEEE Trans Neural Netw.
2003;14(6):1569–72.
3. Brunel N. Dynamics of sparsely connected networks of excitatory and
inhibitory spiking neurons. J Comput Neurosci. 2000;8(3):183–208.
4. Traub RD, Contreras D, Cunningham MO, et al. Single-column thalamocortical network model exhibiting gamma oscillations, sleep spindles,
and epileptogenic bursts. J Neurophysiol. 2005;93(4):2194–2232.
5. Markram H, Muller E, Ramaswamy S, et al. Reconstruction and simulation
of neocortical microcircuitry. Cell. 2015;163(2):456–92.
P61
A synaptic story of dynamical information encoding in neural
adaptation
Luozheng Li1, Wenhao Zhang1, Yuanyuan Mi1, Dahui Wang1,2, Si Wu1
1
State Key Laboratory of Cognitive Neuroscience & Learning, IDG/
McGovern Institute for Brain Research, Beijing Normal University, Beijing
100875, China; 2School of System Science, Beijing Normal University,
Beijing 100875, China
Correspondence: Si Wu ‑ wusi@bnu.edu.cn
BMC Neuroscience 2016, 17(Suppl 1):P61
Adaptation refers to the general phenomenon that a neural system
dynamically adjusts its response property according to the statistics of
external inputs [1]. In response to a prolonged constant stimulation,
neuronal firing rates always first increase dramatically at the onset of
the stimulation; and afterwards, they decrease rapidly to a low level
close to background activity (see Fig. 39A). This attenuation of neural activity seems to be contradictory to our experience that we can
still sense the stimulus after the neural system is adapted [2]. Thus, it
prompts a question: where is the stimulus information encoded during the adaptation? Here, we investigate a computational model in
which the neural system employs a dynamical encoding strategy during the neural adaptation: at the early stage of the adaptation, the
stimulus information is mainly encoded in the strong independent
firings; and as time goes on, the information is shifted into the weak
but concerted responses of neurons (see Fig. 39B). We find that shortterm plasticity [3], a general feature of synapses, provides a natural
mechanism to achieve this goal. Furthermore, we demonstrate that
with balanced excitatory and inhibitory inputs, this correlation-based
information can be read out efficiently. The implications of this study
on our understanding of neural information encoding are discussed.
Page 43 of 112
Fig. 39 Firing rates, synaptic efficacy and cross-correlation change
during the adaptation. A The time course of firing rates and the
averaged synaptic efficacy of the network during the adaptation. ux
is temporally enhanced during the adaptation due to the STF, but in
the long term, strong STD drives the synaptic efficacy to background
level. Stimulation is during 0–1500 ms. B The enhancement of crosscorrelation between neurons during the adaptation
Conclusions We have explored a dynamical encoding strategy in
neural adaptation. By constructing a computational model, we show
that this can be achieved through varying the information encoder
during the adaptation, that is, at the early stage of the adaptation, the
stimulus information is mainly encoded in the strong and independent firings of neurons; and as time goes on, the stimulus information is
shifted into the weak but concerted responses of neurons. This shift of
information encoder can be naturally implemented via STP, a general
feature of synapses.
References
1. Wark B, Lundstrom B N, Fairhall A. Sensory adaptation. Curr Opin Neurobiol. 2007;17(4):423–9.
2. Christopher deCharms R, Merzenich MM. Primary cortical representation of sounds by the coordination of action-potential timing. Nature.
1996;381:13.
3. Markram H, Wang Y, Tsodyks M. Differential signaling via the same axon of
neocortical pyramidal neurons. Proc Natl Acad Sci. 1998;95(9):5323–28.
P62
Physical modeling of rule‑observant rodent behavior
Youngjo Song1, Sol Park1,2, Ilhwan Choi2, Jaeseung Jeong1, Hee‑sup Shin2
1
Bio and Brain Engineering, KAIST, Daejeon, 34141, Republic of Korea;
2
Center for Cognition and Sociality, IBS, Daejeon, 34047, Republic of Korea
Correspondence: Jaeseung Jeong ‑ jsjeong@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P62
There is training room for a mouse. In the room, there are two lights
which is used as a cue for reward, and there are two reward zone.
If a mouse go left reward zone when the left light cue turns on, the
mouse gets reward, and if a mouse go right reward zone when the
right light cue turns on, the mouse gets reward. The reward is given
by brain stimulation from the electrode implanted in MFB. A pair of
mice trained individually in the room in order to make them understand the meaning of two light cues. After individual training, the
two mice released in the same training room at the same time. In this
experiment, 15 out of 19 pairs showed tendency to separate their own
reward zone.
In order to explain the rodent behavior, we made a computational
rodent model. This model is based on Rescorla–Wagner Model. We
defined a success rate as probability that a mouse is in the correct
reward zone when any cue is given. In each trial, the model mouse
learns the left success rate and the right success rate by reinforcement
learning. In this model, we assume that success rates for the left cue
and right cue are independent and we eliminate social interaction
between two model mice.
Results The simulation result is given in below figures. Figure 40A is a
graph of success rate of a pair of model mice which shows rule between
them. You can see the right que success rate of mouse1 and the left que
success rate of mouse2 converge to 1, but the left que success rate of
mouse1 and the right que success rate of mouse2 converge below 0.4.
It means that mouse1 tends to move only when the right cue is given,
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 40 A Success rate of two model rats which shows rule between
them (blue dots represent the left cue success rate of model rat1,
orange plus represent the right cue success rate of model rat1, yellow
cross represent the left cue success rate of model rat2, and purple
line represent the right cue success rate of model rat2). B Simulation
result (600 iterations). C Behavior experiment result (19 pairs)
and mouse2 tends to move only when the left cue is given. Our model
mice, however, shows this rule with less probability than actual behavior result. Figure 40B, C 43 % of model mice showed rule, but 79 % of
actual mice showed rule in behavior experiment.
Conclusion If we assume the mouse as a simple independent creature
which doesn’t have social element, they show rules with lower probability than the actual mouse in behavior experiment. It means that
we can’t regard the mouse as a non-social creature. Therefore, we have
to add the social factors such as empathy or cooperation to simulate
actual rodent behavior.
Reference
1. Glimcher PW, Fehr E. Neuroeconomics, 2nd ed. Academic Press.
P64
Predictive coding in area V4
and prefrontal cortex explains dynamic discrimination of partially
occluded shapes
Hannah Choi1,2,3, Anitha Pasupathy2,3, Eric Shea‑Brown1,3
1
Department of Applied Mathematics, University of Washington,
Seattle, WA 98195, USA; 2Department of Biological Structure,
University of Washington, Seattle, WA 98195, USA; 3UW Institute
for Neuroengineering, University of Washington, Seattle, WA 98195, USA
Correspondence: Hannah Choi ‑ hannahch@uw.edu
BMC Neuroscience 2016, 17(Suppl 1):P64
The visual system recognizes objects in natural scenes without difficulty, even when most objects are partially occluded. The neural
basis of this capacity is unknown. Recent results from primate area
V4, an intermediate stage in the shape processing pathway, suggest
that feedback from higher cortices may be important for the emergence of V4 shape selective signals [1] when animals are engaged in
discriminating partially occluded shapes. Here we implement predictive coding, which has been previously applied to explain responses
in early visual areas [2], to investigate possible mechanisms underlying robust discrimination of partially occluded shapes in V4. We
propose that higher cortical areas such as prefrontal cortex (PFC)
make predictions about V4 activities; when these PFC signals are
relayed via feedback to V4, they can reproduce the delayed peak
of V4 responses observed in experiments. With a model (Fig. 41A)
composed of PFC and V4 units that are selective for different input
features, we capture response characteristics of V4 and PFC measured in experiments, by combining feed-forward sensory inputs
and feedback predictions to maximize the posterior probability of
the responses. We found that inclusion of the feedback predictions
results in stronger shape-selective responses across a range of occlusion levels (Fig. 41B), thus maintaining robust discrimination of partially occluded shapes (Fig. 41C).
Page 44 of 112
Fig. 41 A Schematic of the V4-PFC network model. B Optimal
representation of the shape-selective V4 responses as a function of
occlusion level. C Neuronal responses with a noise projected onto the
test shape-selective (unit 1)/non-selective (unit 2) V4 response plane,
before (top) and after (bottom) the feedback inputs from PFC. Feedback inputs move the responses away from the unity line, improving
shape discriminability under occlusion
Acknowledgements: This research was supported by the Washington
Research Foundation Innovation Postdoctoral Fellowship in Neuroengineering (HC), National Science Foundation CRCNS Grant IIS1309725
(AP), and NEI Grant R01EY018839 (AP).
References
1. Kosai Y, El-Shamayleh Y, Fyall AM, Pasupathy A. The role of visual area
V4 in the discrimination of partially occluded shapes. J Neurosci.
2014;34:8570–84.
2. Rao RPN, Ballard DH. Predictive coding in the visual cortex: a functional
interpretation of some extra-classical receptive-field effects. Nature
Neurosci. 1999;2:79–87.
P65
Stability of FORCE learning on spiking and rate‑based networks
Dongsung Huh1, Terrence J. Sejnowski1,2
1
The Salk Institute for Biological Studies, La Jolla, CA 92037 USA; 2Division
of Biological Sciences, University of California at San Diego, La Jolla, CA
92095 USA
Correspondence: Dongsung Huh ‑ huh@salk.edu
BMC Neuroscience 2016, 17(Suppl 1):P65
Neurons in the brain often exhibit complex activity patterns, with
fluctuations on time scales of several seconds. The generation of complex patterns is critical for directing movements, and is likely to be
involved in processing time-varying input (such as speech). However,
it is not yet understood how networks of spiking neurons, with time
constants of only a few milliseconds, could exhibit such slow dynamics. This should be contrasted with rate-based neural networks, which
can be easily trained to generate arbitrary complex activity patterns in
a reservoir-based manner [1] by an iterative training method (FORCE
learning [2]). So far, however, FORCE learning has not led to successful
training of spiking neural networks.
Here, we analyze the stability of the networks that result from such
learning schemes. For linear rate-based networks, we can analytically
predict the full dynamic property of the networks. As the network’s
recurrent connectivity reaches the “edge of chaos”, the neuronal
activity exhibits a broad distribution of phase, providing appropriate
basis for generating the fluctuations. For weaker recurrent connectivity, however, the phase distribution becomes much narrower. In this
case, the trained network exhibits highly non-normal structure, which
becomes unstable even under small perturbations. Our analysis also
illuminates the source of instability in training spiking networks,
which is mainly due to the rectified nature of the neuronal output.
In numerical simulations, rectified-linear rate networks exhibit narrow phase distribution, even with strong recurrent connectivity near
the edge of chaos. Moreover, introducing spiking-dynamics further
BMC Neurosci 2016, 17(Suppl 1):54
reduces the width of the distribution, leading to highly unstable
network dynamics. Our result reveals the limitation of the reservoirbased approaches, and may lead to more stable, alternative training
methods.
Acknowledgements: Supported by HHMI.
References
1. Jaeger H, Haas H. Harnessing nonlinearity: predicting chaotic systems
and saving energy in wireless communication. Science. 2004;304:78–80.
2. Sussillo D, Abbott LF. Generating coherent patterns of activity from
chaotic neural networks. Neuron. 2009;63(4):544–57.
P66
Stabilising STDP in striatal neurons for reliable fast state
recognition in noisy environments
Simon M. Vogt1, Arvind Kumar2,3, Robert Schmidt1,2
1
BrainLinks‑BrainTools, Cluster of Excellence, University of Freiburg,
Germany; 2Faculty of Biology and Bernstein Center Freiburg, University
of Freiburg, Germany; 3Department of Computational Biology, Royal
Institute of Technology Stockholm, Sweden
Correspondence: Simon M. Vogt ‑ simonsunimail@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P66
The brain must be able to quickly identify environmental states based
on sensory inputs and select appropriate actions. To obtain neurons
that respond selectively to states with short response latencies, a neural plasticity rule is required. Spike timing dependent plasticity (STDP) is
known to find the earliest predictors of spatiotemporal spike patterns
without supervision. While this feature of STDP is often seen as a hindrance when aiming to reproduce an exact target output spike train,
we exploit it to generate short-latency, reliable pattern recognition.
There are, however, some difficulties in using STDP in real-world
continuous scenarios. In models, presynaptic firing rates are often
assumed to be stationary or have constant correlation for simplification, and STDP rule parameters must be closely fitted to form a
stable fixed point in the postsynaptic firing rate and normalise postsynaptic activity. Any deviation from these requirements can cause
the postsynaptic neuron to become quiet before it is able to form
strong selectivity, or exhibit a runaway effect that makes the neuron
responsive to a large set of inputs. Soft-bound STDP with a weightdependent attractor has been suggested as a means for stabilising
postsynaptic activity, but this actively hinders separation of spatiotemporal patterns from background noise: as during learning the
synaptic weight moves away from the attractor, noisy input becomes
more likely to undo these weight changes. Furthermore, activitydependent scaling also aims to keep the postsynaptic neuron active
indefinitely, so neurons may lose any learned selectivity when
enduring long periods of background noise. Interpolations between
different types of STDP have been suggested, but a problem of balance between premature selectivity and longtime noise robustness
remains.
We solve this dilemma by including a slow continuous potentiation in our model, which depends on metabolic cost of maintaining
strong synapses and slowly vanishes as neurons become selective. It
is independent of pre- or postsynaptic activity and can recover silent
neurons if a metabolic cost function allows it to. Together with negative-integral STDP as taken from experimental data, it stabilises the
activity of untrained neurons for a much wider range of rule parameters and heterogeneous input activity. Our model maintains a unimodal weight distribution while the postsynaptic neuron has not yet
become selective, but does not impair the formation of selectivity
to spatiotemporal patterns. Selectivity is quickly achieved as soon
as patterns are present, even after enduring long periods of noise.
Connections that only present noise, represent only other patterns,
or present only late parts of a trained pattern become ineffective as
the neuron becomes selective, and may be pruned. Any selectivity
is hence ensured to represent actual spatiotemporal spike patterns
Page 45 of 112
that were at some point present in the postsynaptic neuron’s inputs.
This makes the process of training neurons to detect environmental
states encoded as spatiotemporal patterns more robust to variations
in input statistics and rule parameters, thus easing application in
larger-scale networks.
Our model of fast pattern detection may apply specifically to the striatum of the basal ganglia where fast reliable decisions need to be made
within milliseconds. Unsupervised learning should coexist with rare
dopaminergic reinforcement to continuously form new representations of environmental events and decide which of these events are
behaviourally important and which can safely be ignored.
Acknowledgements: Cluster of Excellence BrainLinks-BrainTools
funded by German Research Foundation (DFG, Grant Number EXC
1086).
P67
Electrodiffusion in one‑ and two‑compartment neuron models
for characterizing cellular effects of electrical stimulation
Stephen Van Wert1, Steven J. Schiff1,2
1
Center for Neural Engineering, Department of Engineering Science
and Mechanics, The Pennsylvania State University, University Park, PA
16802, USA; 2Departments of Neurosurgery and Physics, The Pennsylvania
State University, University Park, PA 16802, USA
Correspondence: Stephen Van Wert ‑ szv124@psu.edu
BMC Neuroscience 2016, 17(Suppl 1):P67
Standard approaches for modeling the neuronal effects of electrical fields and currents (such as [1, 2]) apply transmembrane current
to Hodgkin–Huxley membrane patches without regard to ion fluxes
and conservation of ions inside and outside the cell. We propose cellular models that reflect polarization and preserve the biophysics of
the spaces the neurons are embedded within. By including ion fluxes
and maintaining conservation of mass and charge, the gradients of
ionic concentrations both within and outside of the neuron can be
accounted for. This requires characterizing the ionic fluxes with electrodiffusion, such that ionic charge gradients as well as ionic concentration gradients drive flux. This electrodiffusion mechanism, derived
from Nernst-Planck flux equation, not only allows for more accurate
modeling of physiological and pathophysiological conditions with
substantial ionic and volume changes, but it also provides a means
to model application of electrical stimulation. The model being developed here builds on recent one-compartment model development
in [3] that extends the Hodgkin–Huxley formalism in several distinct
ways, most notably in using conservation to track all ion fluxes and
volume changes to determine the extra- and intra-cellular concentrations. The proposed model also extends the model in [3] to a two
compartment model which allows for simulation of neuronal polarization with control applied in the direction of a soma-dendritic axis.
We first characterize the resultant model dynamics in the absence
of any control stimulus and compare these to the dynamics seen
with standard diffusion. In particular, the dynamics are characterized
through trajectories of dynamically evolving variables, with a focus on
bifurcation structure at points where the dynamics transition from different states such as normal firing, seizure, or spreading depression.
We then simulate the effects of applying excitatory or inhibitory control on these dynamics and optimize the dynamics of the neuron to
be consistent with experimental evidence. Such a model gives very
different results from the customary approach to modeling the effects
of electrical stimulation, where the stimulation is applied internally or
externally to a neuron without taking the nature of the charge carriers present into account. There are a variety of effects of excitatory and
inhibitory stimulation observed now that could not be possible before,
and we can describe the trajectories of the effects of such stimulation
in ways that shed light on multiple experimental scenarios.
This type of model development offers the ability to understand a
wide variety of previously unexplained experimental observations for
both excitatory and inhibitory stimulation. Doing so offers a platform
BMC Neurosci 2016, 17(Suppl 1):54
for us to study electrical feedback control of neuronal systems and to
offer model-based control strategies for pathological dynamics such
as seizures and spreading depression.
References
1. Park E-H, Barreto E, Gluckman BJ, Schiff SJ, So P. A model of the effects of
applied electric fields on neuronal synchronization. J Comput Neurosci.
2005;19: 53–70.
2. Berzhanskaya J, Chernyy N, Gluckman BJ, Schiff SJ, Ascoli GA. Modulation
of hippocampal rhythms by subthreshold electric fields and network
topology. J Comput Neurosci. 2013;34(3):369–89.
3. Wei Y, Ullah G, Schiff SJ. Unification of neuronal spikes, seizures, and
spreading depression. J Neurosci. 2014;34(35):11733–43.
P68
STDP improves speech recognition capabilities in spiking
recurrent circuits parameterized via differential evolution Markov
chain Monte Carlo
Richard Veale1, Matthias Scheutz2
1
National Institute for Physiological Sciences, Okazaki, Aichi, Japan;
2
Department of Computer Science, Tufts University, Medford, MA, USA
Correspondence: Richard Veale ‑ richard@nips.ac.jp
BMC Neuroscience 2016, 17(Suppl 1):P68
A major issue in using spiking neural circuits for pragmatic tasks such
as speech recognition is how to parameterize them. Here, we apply a
hybrid Differential Evolution/Markov Chain Monte Carlo (DE/MCMC) [1,
2] approach to estimate optimal parameters for a spiking neural circuit
that is used for real time speech recognition [3] from raw auditory input
using PSWEEP2 (rveale.com/software.php). To avoid the expensive
training step, we use a surrogate measure of word recognition performance. Specifically, we maximize average within-word similarity in the
neural circuit’s state space trajectory, while simultaneously minimizing
between-word similarity. We executed the algorithm for 7000 generations (48 h of runtime) using 2016 cores of the super computer Big Red
II at Indiana university. The average fitness increases significantly with
successive generations. Panel a is a visualization of the first 3 principle
components of the state space for the exemplars of each different word
category (shown as different colors). The different categories move to
take more distant trajectories through the state space with successive
generations. We verify that the state-space separation is a good surrogate measure of word recognition performance by taking the set of
circuits with the highest fitness from the first, middle, and last 100 generations and training readout neurons for the 7-word corpus. Word recognition performance increases from 19 to 71 to 85 % (Fig. 42).
Finally, we evaluated the performance benefit of adaptation to sensory stimuli via synaptic plasticity mechanisms, to make pragmatic
use of our previous work investigating auditory habituation [4]. We
take the most performant parameter points that were found during
the parameter sweep, and test their fitness before and after exposure to 100 presentations of the word stimuli while a nearest-neighbor temporally asymmetric Hebbian plasticity model of spike timing
dependent plasticity (STDP) is implemented in all excitatory synapses.
Although sensitive to STDP model parameters, Panel B shows that
word recognition performance can be improved by as much as 8 % by
familiarizing a neural circuit to the type of sensory stimulus that it will
be used to compute. This follows previous findings by Triesch et al. [5],
who reported similar effects in non-spiking neural networks.
Fig. 42 A Fitness evolution generation 0–4000 (first 3 principle
components). Each color is a different word class, each line a different
utterance token of the word. B Change in state space trajectory from
STDP adaptation
Page 46 of 112
References
1. Veale R, Isa T, Yoshida M. Applying differential evolution MCMC to
parameterize large-scale spiking neural simulations. In: IEEE conference
on evolutionary computation (CEC). IEEE. p. 1620–27.
2. Laloy E, Vrugt JA. High-dimensional posterior exploration of hydrologic
models using multiple-try dream (zs) and high-performance computing,
Water Resour Res. 2012;48(1).
3. Veale R, Scheutz M. Neural circuits for any-time phrase recognition. In:
Proceedings of the 34th annual conference of the Cognitive Science
Society; 2012. p. 1072–7.
4. Veale R, Scheutz M. Auditory habituation via spike-timing dependent. In:
Proceedings of the international conference on development and learning and epigenetic robotics (ICDL), San Diego, CA; 2012.
5. Lazar A, Pipa G, Triesch J. Fading memory and time series prediction
in recurrent networks with different forms of plasticity. Neural Netw.
2007;20(3):312–22.
P69
Bidirectional transformation between dominant cortical neural
activities and phase difference distributions
Sang Wan Lee1,2,3
1
Department of Bio and Brain Engineering, Korea Advanced Institute
of Science and Technology (KAIST), Daejeon, South Korea; 2Program
of Brain and Cognitive Engineering, Korea Advanced Institute of Science
and Technology (KAIST), Daejeon, South Korea; 3KAIST Institute
for Health Science and Technology, Korea Advanced Institute of Science
and Technology (KAIST), Daejeon, South Korea
Correspondence: Sang Wan Lee ‑ sangwan@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P69
The brain is a complex nonlinear dynamic system comprising multiple
different types of subsystems. Each subsystem encodes different types
of information, and its states are context-dependent. Unlike sensory
or motor processing occurring at relatively early or terminal stages
of functional hierarchy, cognitive processes, including learning, inference, and top-down attention, require interactions between brain’s
multiple subsystems. The associated neural dynamics inevitably leave
an imprint on neural activity patterns over a wide areas of cortex.
Considerable progress has been made toward understanding such functional network dynamics. This includes the causal connectivity [1], its
extension to distinguish causality from correlation within nonseparable
weakly connected dynamic systems [2], and the integrated information
theory to quantify the effect of a neuronal network connectivity on the
increase in the amount of information above and beyond the capability of
a single locally connected network [3, 4]. However, none of these methods
is applicable to real-time analyses when the network size is large.
Here I develop a simple and efficient computational framework for
analyzing cortical dynamics both in time and space, arising from
complex interactions between brain’s multiple subsystems. Accommodating the fact that both a covariance and a gram matrix can be
computed by using a combination of a certain idempotent matrix with
a data matrix, I derive a set of matrix operators F, with which one set
of eigenvectors associated with a covariance matrix of data and of
mean-corrected data can be transposed to another set of eigenvectors associated with a gram matrix, and vice versa (see Fig. 43). For
example, suppose that a d-by-n data matrix is a set of time-series data
recorded from multiple locations of cortices where d and n refers to
the number of electrodes and time points, respectively, and d ≪ n.
One can then perform a singular value decomposition (SVD) for the
d-by-d covariance or gram matrix to obtain an associated eigenvector
set, followed by applying matrix operator F′ to the eigenvector set to
convert it to the eigenvector set of their counterparts. There is no need
to perform SVD for n-by-n matrix whose computational load is high. If
n ≪ d, then one could simply start with performing the SVD for the
n-by-n matrix, followed by applying the matrix operator F to its eigenvector set. It is noted that the acquired d-by-1k eigenvectors and n-by1k eigenvectors correspond to dominant cortical neural activities and
phase difference distributions, respectively.
Conclusion An efficient computational framework for analyzing cortical dynamics both in time and space is proposed, taking into account
the relationship between a covariance and a gram matrix. For analyzing
BMC Neurosci 2016, 17(Suppl 1):54
Page 47 of 112
Fig. 43 Computational framework for analyzing space–time cortical
dynamics. T is an idempotent projection matrix
neural data acquired from multiple locations of cortices, the framework replaces the SVD with a simple matrix operator F so as to reduce a
heavy computational load of performing SVD on large-size data matrices. In doing so, it allows efficient bidirectional transformation between
dominant neural activities and phase difference distributions.
Acknowledgements: I thank Barclay Lee and Dae-Hyun Kim for their
assistance. This work was supported by the research fund of the KAIST
(Korea Advanced Institute of Science and Technology) (Grant code:
G04150045).
References
1. Seth AK. Causal connectivity of evolved neural networks during behavior.
Network. 2005;16:35–54.
2. Sugihara G, May R, Ye H, Hsieh C, Deyle E, Fogarty M, Munch S. Detecting
causality in complex ecosystems. Science. 2012;338(6106):496–500.
3. Balduzzi D, Tononi G. Qualia: the geometry of integrated information.
PLoS Comput Biol. 2009;5(8):e1000462.
4. Edlund JA, Chaumont N, Hintze A, Koch C, Tononi G, Adami C. Integrated
information increases with fitness in the evolution of animats. PLoS
Comput Biol. 2011;7(10):e1002236.
P70
Maturation of sensory networks through homeostatic structural
plasticity
Júlia Gallinaro1, Stefan Rotter1
1
Bernstein Center Freiburg and Faculty of Biology, University of Freiburg,
Freiburg, Baden‑Württember, 79194, Germany
Correspondence: Simon M. Vogt ‑ julia.gallinaro@bcf.uni‑freiburg.de
BMC Neuroscience 2016, 17(Suppl 1):P70
Neurons in the adult visual cortex of mice prefer to make synapses
with neurons responding to similar visual features. As such a bias in
connectivity is not observed at the time of eye opening, it has been
proposed that the functional subnetworks are formed through rewiring of recurrent synaptic connections, induced by visual experience
[1]. However, it is not clear according to which rules this structure
develops. The emergence of feature specific wiring was recently demonstrated in a balanced network model with appropriate rules of functional synaptic plasticity [2]. In this model, however, connectivity was
evaluated based on the strength of already existing synapses, and
the structure of the network remained unchanged throughout the
simulation.
Referring to recent findings of homeostatic regulation of cortical activity
in rodent visual cortex in vivo [3], we employ here a structural plasticity
rule based on firing rate homeostasis described previously [4] for simulating network restructuring during sensory stimulation. We show that,
next to other biologically meaningful properties, feature specific connectivity also emerges in a balanced network of changing structure (see
Fig. 44), using a plasticity rule that does not depend on spike timing.
Fig. 44 Network connectivity before and after sensory stimulation.
A, B. Connectivity matrix, pre- and post-synaptic neurons are sorted
according to their preferred orientation (PO) and subdivided into
groups. C, D. Mean output connectivity plotted against the difference
between pre and post PO
Acknowledgements: Supported by the Erasmus Mundus Joint Doctoral program EuroSPIN, the German Federal Ministry of Education
and Research, grant 01GQ0830, and the state of Baden-Württemberg
through bwHPC.
References
1. Ko H, Cossell L, Baragli C, Antolik J, Clopath C, Hofer SB, Mrsic-Flogel
TD: The emergence of functional microcircuits in visual cortex. Nature.
2013;496:96–100.
2. Sadeh S, Clopath C, Rotter S. Emergence of functionalspecificity in balanced networks with synapticplasticity. PLoS Comput Biol.
2015;11:e1004307.
3. Bishop HI, Zito K. The downs and ups of sensorydeprivation: evidence for
firing rate homeostasis in vivo. Neuron. 2013;80:247–9.
4. van Ooyen A: Using theoreticalmodels to analyseneuraldevelopment. Nat
Rev Neurosci. 2011;12:311–26.
P71
Corticothalamic dynamics: structure, number of solutions
and stability of steady‑state solutions in the space of synaptic
couplings
Paula Sanz‑Leon1,2, Peter A Robinson1,2
1
School of Physics, University of Sydney, New South Wales, Australia;
2
Center for Integrative Brain Function, University of Sydney, New South
Wales, Australia
Correspondence: Paula Sanz‑Leon ‑ paula.sanz‑leon@sydney.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P71
The interconnections of a model of the corticothalamic system [1]
define an 8-dimensional parameter space where specific combinations of dimensions correspond to one of the three loops of the system (e.g., intracortical, corticothalamic and intrathalamic). The form of
the steady-state equation of the corticothalamic system imposes an
odd number of solutions, which in terms of dynamics correspond to
fixed points of the system. Here, the structure of regions with different
number of solutions is systematically investigated within physiologically valid ranges of synaptic couplings representing different brain
states [2, 3]. For instance, Fig. 45A, Bdisplay the regions where the
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 45 Three dimensional subsets of the 8D corticothalmic coupling
space. A, B Regions with 1, 3 or 5 roots are enclosed by surfaces (blue,
violet and yellow respectively). The sharp transition between zones
of 1–3 roots along the vee axis (excitatory intracortical feedback)
indicates the plane at which the total intracortical feedback (vee + vei)
changes sign. The difference between A, B is the value of vei (inhibitory intracortical feedback). In a the probability of having multiple
roots is lower than in B
steady state equation has one, three or five solutions for two 3-dimensional subsets of the full space. These results show how small changes
in the connectivity can cause additional roots of the steady state equation to appear or vanish. More importantly, they illustrate the effect
of intracortical feedback: for more than one solution to exist the total
intracortical feedback needs to be negative (inhibitory). The occurrence of multiple roots happens for parameter values that characterize
normal arousal states [3], indicating that the approach presented here
has a potential to (i) quantify and predict the existence of additional
(abnormal) arousal states and (ii) categorize subtle differences in states
such as anesthesia, coma [4].
References
1. Robinson PA, Rennie CJ, Wright JJ, Bourke PD. Steady states and global
dynamics of electrical activity in the cerebral cortex. Phys Rev E.
1998;58:3557–71.
2. Robinson PA, Rennie CJ, Wright JJ, Bahramali H, Gordon E, Rowe DL. Prediction of electroencephalographic spectra from neurophysiology. Phys
Rev E. 2001; 63:021903.
3. Abeysuriya RG, Rennie CJ, Robinson PA. Physiologically based arousal
state estimation and dynamics. J Neurosci Methods. 2015;253:55–69.
4. Steyn-Ross ML, Steyn-Ross DA, Sleigh JW, Wilcocks LC. Toward a theory of
the general-anesthetic-induced phase transition of the cerebral cortex. I.
A thermodynamics analogy. Phys Rev E. 2001;64:942–53.
P72
Optogenetic versus electrical stimulation of the parkinsonian
basal ganglia. Computational study
Leonid L. Rubchinsky1,2, Chung Ching Cheung1, Shivakeshavan
Ratnadurai‑Giridharan1
1
Department of Mathematical Sciences, Indiana University‑Purdue
University Indianapolis, Indianapolis, IN, USA; 2Stark Neurosciences
Research Institute, Indiana University School of Medicine, Indianapolis,
IN, USA
Correspondence: Leonid L. Rubchinsky ‑ lrubchin@iupui.edu
BMC Neuroscience 2016, 17(Suppl 1):P72
Deep brain stimulation (DBS) is used as a therapeutic procedure to treat
symptoms of several neurological and neuropsychiatric disorders. In
particular it is used to treat motor symptoms of Parkinson’s disease (PD)
by delivering high-frequency regular stimulation to subcortical targets.
Hypokinetic symptoms of PD are associated with excessive oscillatory
synchronized activity in the beta frequency band, and effective DBS is
believed to suppress it. An alternative way to stimulate neural circuits
is an emerging technology of optogenetics. It is an experimental technique and it is not clear if it eventually will be possible to implement
it in clinical practice. However it is used as an experimental tool, and
maybe, in time, it will be developed into safe therapeutic technique.
Page 48 of 112
The goal of his study is to explore how effective an optogenetic
stimulation in comparison with electrical stimulation in their network
effects on elevated synchronized oscillatory activity. We use a model
for the basal ganglia activity [1], which was developed to reproduce
experimentally observed beta-band activity patterns [2]. We introduce electrical stimulation as well as optogenetic stimulation of two
types: excitatory via channelrhodopsin and inhibitory via halorodopsin. We explore the effect of different stimulation types on oscillatory
synchronized dynamics and consider the efficacy of stimulation for
different kind of network’s dynamics.
All three modes of stimulation can decrease beta synchrony that is
commonly associated with hypokinetic symptoms of Parkinson’s disease. Generally speaking, growing intensity of stimulation leads to
larger suppression of the beta-band synchronized oscillatory activity. But the actions of different stimulation types on the beta activity
may differ from each other. Electrical DBS and optogenetic excitation
have somewhat similar effects on the network. Both of these stimulation types cause desynchronization and suppression of the beta-band
bursting. As intensity of stimulation is growing, they synchronize the
network at higher (non-beta) frequencies in a close to tonic spiking dynamics. Optogenetic inhibition effectively reduces spiking and
bursting activity of the targeted neurons.
We compare the stimulation modes in terms of the minimal effective
current delivered to basal ganglia neurons in order to suppress beta
activity below a threshold: the less stimulation current is needed to
suppress the activity, the more efficacious stimulation is. We found
that optogenetic inhibition usually requires less effective current than
electrical DBS to achieve beta suppression. Optogenetic excitation,
while as not efficacious as optogenetic inhibition, still usually requires
less effective current than electrical DBS to suppress beta activity.
Thus our results suggest that optogenetic stimulation may introduce
less of effective currents to a neuron than conventional electrical DBS,
but still achieve sufficient beta activity suppression. Optogenetics is
presently not used in humans. However, it was implemented in the
basal ganglia of non-human primates [3]. So we suppose our results
may motivate further research into applicability of optogenetic technologies in humans. Optogenetic stimulation is also used as a research
tool. Our results suggest that it may be more effective than electrical
stimulation in control of synchronized oscillatory activity, because it
does its job with less current injected into the neurons.
Acknowledgements: The study was supported by ICTSI and the Indiana University Health – Indiana University School of Medicine Strategic Research Initiative.
References
1. Park C, Worth RM, Rubchinsky LL. Neural activity in Parkinsonian brain: the
boundary between synchronized and nonsynchronized dynamics. Phys
Rev E. 2011;83:042901.
2. Park C, Worth RM, Rubchinsky LL. Fine temporal structure of beta oscillations synchronization in subthalamic nucleus in Parkinson’s disease. J
Neurophysiol. 2010;103:2707–16.
3. Galvan A, Hu X, Smith Y, Wichmann T. In vivo optogenetic control
of striatal and thalamic neurons in non-human primates. PLoS One.
2012;7(11):e50808.
P73
Exact spike‑timing distribution reveals higher‑order interactions
of neurons
Safura Rashid Shomali1, Majid Nili Ahmadabadi1,2, Hideaki Shimazaki3, S
Nader Rasuli4,5
1
School of Cognitive Sciences, Institute for Research in Fundamental
Sciences (IPM), Tehran, 19395‑5746, Iran; 2School of ECE, College
of Engineering, University of Tehran, Tehran, 14155‑6619, Iran; 3RIKEN
Brain Science Institute, Wako, Saitama, 351‑0198, Japan; 4Department
of Physics, University of Guilan, Rasht, 41335‑1914, Iran; 5School
of Physics, Institute for Research in Fundamental Sciences (IPM), Tehran,
19395‑5531, Iran
Correspondence: Safura Rashid Shomali ‑ safura@ipm.ir
BMC Neuroscience 2016, 17(Suppl 1):P73
BMC Neurosci 2016, 17(Suppl 1):54
It has been suggested that variability in spike patterns of individual
neuron is largely due to noisy fluctuations caused by asynchronous
synaptic inputs balanced near the threshold regime [1–3]. In this
regime, small fluctuations in synaptic inputs to a neuron do cause
output spikes; because the membrane potential is maintained below
but close enough to the threshold potential. To successfully transfer
signals under such noisy conditions, it is proposed that a few relatively
stronger synapses and/or an assembly of nearly synchronous ones
form “signaling inputs” [4]. Thus one fundamental question is how
such relatively strong signaling input modifies the spiking activity of
a post-synaptic neuron which receives noisy background inputs balanced near the threshold regime. Nonetheless, analytical studies on
the effect of the signaling input under such conditions are scarce even
with the popular leaky integrate-and-fire (LIF) neuron model. Here we
analytically study the impact of a specified signaling input on spike
timing of the postsynaptic LIF neuron which receives noisy inputs at
the threshold regime. To this end, we first revisit Fokker–Planck analysis of a first spike-timing distribution when the LIF neuron receives
noisy synaptic inputs, but no signaling input, at the threshold regime.
We then perform perturbation analysis to investigate how a signaling input modifies this first spike-timing distribution. Fortunately, we
could solve all terms of perturbation analytically and find the exact
first spike-timing distribution of the postsynaptic neuron; it is applicable to not only excitatory but also inhibitory input. This analytical
solution allows us to describe the statistics of output spiking activity
as a function of background noise, membrane dynamics, and signaling
input’s timing and amplitude.
The proposed analysis of signaling input provides a powerful framework for studying information transmission, neural correlation, and
timing-dependent synaptic plasticity. Among them, we investigate
the impact of common signaling inputs on population activities of
postsynaptic neurons. Using mixture models based on our analytical
first spike-timing distribution, we calculate the higher-order interactions [5] of postsynaptic neurons in different network architectures.
Comparing these results with higher-order interactions, measured
from experimental data in monkey V1 [6], we try to answer whether
one can reveal network architecture, responsible for the ubiquitously
observed sparse activities.
References
1. Vreeswijk C, Sompolinsky H. Chaos in neuronal networks with balanced
excitatory and inhibitory activity. Science. 1996;274: 1724–6.
2. Renart A, De La Rocha J, Bartho P, Hollender L, Parga N, Reyes A, Harris KD.
The asynchronous state in cortical circuits. Science. 2010;327:587–90.
3. Tan AY, Chen Y, Scholl B, Seidemann E, Priebe NJ. Sensory stimulation
shifts visual cortex from synchronous to asynchronous states. Nature.
2014;509:226–9.
4. Teramae Jn, Tsubo Y, Fukai T. Optimal spike-based communication in
excitable networks with strong-sparse and weak-dense links. Sci Rep.
2:485.
5. Nakahara H, Amari S. Information-geometric measure for neural spikes.
Neural Comput. 2002;14:2269–316.
6. Ohiorhenuan IE, Victor JD. Information-geometric measure of 3-neuron
firing patterns characterizes scale dependence in cortical networks. J
Comput Neurosci. 2011;30:125–41.
P74
Neural mechanism of visual perceptual learning using a
multi‑layered neural network
Xiaochen Zhao1, Malte J. Rasch1
1
State Key Lab of Cognitive Neuroscience and Learning, IDG/McGovern
Institute for Brain Research, Beijing Normal University, Beijing 100875,
China
Correspondence: Malte J. Rasch - malte.rasch@bnu.edu.cn
BMC Neuroscience 2016, 17(Suppl 1):P74
Recently, a study [1] has found by recording the activities of neurons
in monkeys performing a contour detection task that the response
properties of the primary visual cortex (V1) change continuously
Page 49 of 112
during perceptual learning. In particular, the figure-background contrast was continuously enhanced in the course of learning. However,
the exact neural circuit mechanisms that causes the V1 responses
to change during perceptual learning remain unclear. In order to
understand how the underlying neural network needs to change, we
here train a multi-layered neural network model to perform the contour detection task on the very same visual stimuli as in the experiments and investigate the network’s performance and the resulting
synaptic weight structure.
In this study, we first model the V1 representation of each visual
stimulus by using a non-classical receptive field model (NCRF) which
takes into account orientation selective inhibition [3]. We further
assume that the higher visual areas (up to a decision unit) are hierarchically structured, read out the V1 activity, and learn to change their
synaptic weights to optimally perform the contour detection task.
AGREL (attention-gated reinforcement learning) algorithm [2], which
considers feedback connections and biologically plausible local synaptic adjustments, is applied to train the network (see Fig. 46). We
found that the multi-layered model trained with AGREL could replicate the behavioral performance increase in a contour detection task
as observed in experiments. Moreover, learning the network model
structure showed enhanced synaptic weights in the region of the
detected contour. It further demonstrated that “predictive” feedback
signals from higher layers facilitate the responses of V1 neurons to
the contour and thus increased the figure-background contrast in
V1 with improved behavioral performance. The results suggest that
the experimental observed V1 response facilitation could be caused
by selective synaptic strengthening of feed-forward and feed-back
pathways.
References
1. Yan Y, Rasch MJ, Chen M, et al. Perceptual training continuously refines
neuronal population codes in primary visual cortex. Nat Neurosci.
2014;17(10):1380–7.
2. Roelfsema PR, van Ooyen A. Attention-gated reinforcement learning of internal representations for classification. Neural Comput.
2005;17(10):2176–214.
3. Zeng C, Li Y, Yang K, et al. Contour detection based on a non-classical
receptive field model with butterfly-shaped inhibition subregions. Neurocomputing. 2011;74(10):1527–34.
Fig. 46 V1 representations of the stimuli and simulation results of
the model. A V1 representations of the stimuli by using NCRF model.
B Model performance increases with perceptual training. C Synaptic
weight changes after perceptual learning
BMC Neurosci 2016, 17(Suppl 1):54
P75
Inferring collective spiking dynamics from mostly unobserved
systems
Jens Wilting1, Viola Priesemann1,2
1
Max‑Planck‑Institute for Dynamics and Self‑Organization, D‑37077
Göttingen, Germany; 2Bernstein Center for Computational Neuroscience,
University of Göttingen, D‑37075 Göttingen, Germany
Correspondence: Jens Wilting ‑ jwilting@nld.ds.mpg.de
BMC Neuroscience 2016, 17(Suppl 1):P75
What can we know about a high-dimensional dynamical system if we
can only observe a very small part of it? This problem of spatial subsampling is common to almost every area of research where spatially
extended, time evolving systems are investigated, and is particularly
severe when assessing population spiking dynamics in neuroscience.
Previous studies have shown that subsampling can lead to spurious results when assessing the dynamical state of spiking activity, in
particular when discriminating whether neural networks operates at
criticality [1, 2]. Here we present further insight why the distance to
criticality is systematically overestimated, and introduce a novel estimator which for the first time allows to correctly infer the distance to
criticality even under strong subsampling.
Neuronal systems have been proposed to operate close to criticality,
because in models criticality maximizes information processing capacities [e.g. 3]. Indeed, power-law distributions of the avalanche size, an
indication of criticality, have been found for local field potentials from
in vitro systems [1] to humans in vivo [4]. However, for neuronal systems criticality also comes with the risk of spontaneous runaway activity, which may lead to pathological states like epilepsy. Experiments
indeed indicate that spiking activity in rats, cats, and monkeys is in a
sub-critical regime, thereby keeping a safety-margin from criticality
[5]. Quantifying the precise distance to criticality may help to shed
light on how the brain maximizes its information processing capacities
without risking runaway activity.
In neural systems, critical dynamics is typically compared to dynamics from models that resemble branching processes [1]. Their dynamics are controlled by a single parameter, the expected number σ of
postsynaptic spikes generated by one individual spike, showing either
stationary dynamics (sub-critical, σ < 1) or transient growth (supercritical, σ > 1). For σ = 1 branching processes are critical and produce
heavy tailed avalanche size distributions. We used a driven branching
process, which allows to exactly match the model neuron firing rate to
that observed in experiments for any σ. We propose a stochastic representation of subsampling and show that under subsampling established approaches to inferring σ are substantially biased. We derived
a novel approach based on multistep regression [6], which for the first
time allows to quantify the distance to criticality even under strong
subsampling. Our method generalizes to auto-regressive processes
with both additive and multiplicative noise, making it widely applicable in diverse fields of research. We validate our method by applying
subsampling to simulated branching networks with invasion, and also
to a network of integrate-and-fire neurons.
We applied this method to spike recordings from awake macaque
monkeys prefrontal cortex, cat visual cortex, and rat hippocampus. We
found that neuronal population activity operates close to criticality,
but in a subcritical regime with 0.94 < σ < 0.995. These results point
at a novel universal organization principle: spiking dynamics in vivo is
in a subcritical regime which does not yield maximum, but sufficient
information processing capacity, and at the same time keeps a safetymargin from unstable supercritical states.
References
1. Beggs J, Plenz D. Neuronal avalanches in neocortical circuits. J Neurosci.
2003;23(35):11167–77.
2. Priesemann V, Munk MHJ, Wibral M. Subsampling effects in neuronal
avalanche distributions recorded in vivo. BMC Neurosci. 2009;10:40.
3. Boedecker J, Obst O, Lizier JT, Mayer NM, Asada M. Information processing in echo state networks at the edge of chaos. Theory Biosci.
2012;131:205–13.
Page 50 of 112
4.
5.
6.
7.
Priesemann V, Valerrame M, Wibral M, Le Van Quyen M. Neuronal avalanches differ from wakefulness to deep sleep—evidence from intracranial depth recordings in humans. PloS Comput Biol. 2013;9(3):e1002985.
Arviv O, Goldstein A, Shriki O. Near-critical dynamics in stimulus-evoked
activity of the human brain and its relation to spontaneous resting-state
activity. J Neurosci. 2015;35(41):13927–42.
Priesemann V, et al. Spike avalanches in vivo suggest a driven, slightly
subcritical brain state. Front Syst Neurosci. 2014;8:108.
Wilting J, Priesemann V. Quantifying the distance to criticality under
subsampling. BMC Neurosci. 2015;16:O3.
P76
How to infer distributions in the brain from subsampled
observations
Anna Levina1, Viola Priesemann2
1
IST Austria, Klosterneuburg, 3400, Austria; 2BCCN & MPI for Dynamics
and Self‑Organization, Göttingen, 37077, Germany
Correspondence: Anna Levina ‑ anna.levina@ist.ac.at
BMC Neuroscience 2016, 17(Suppl 1):P76
Inferring the dynamics of a system from observations is a challenge,
even if one can observe all system units or components. The same task
becomes even more challenging if one can sample only a small fraction of the units at a time. As a prominent example, spiking activity in
the brain can be accessed only for a very small fraction of all neurons
in parallel. These limitations do not affect our ability to infer single
neuron properties, but it influences our understanding of the global
network dynamics or connectivity: Subsampling can hamper inferring
whether a system shows scale-free topology or scale-free dynamics
(criticality) [1, 2]. Criticality is a dynamical state that maximizes information processing capacity in models, and therefore is a favorable
candidate state for brain function. Experimental approaches to test
for criticality extract spatio-temporal clusters of spiking activity, called
avalanches, and test whether they followed power laws. Avalanches
can propagate over the entire system, thus observations are strongly
affected by subsampling. We developed a formal ansatz to infer avalanche distributions in the full system from spatial subsampling using
both analytical and numerical approaches.
In the mathematical model subsampling from exponential distribution does not change the class of distribution, but only its parameters.
In contrast, power law distributions, despite their alias “scale-free”, do
not manifest as power laws under subsampling [2]. We study changes
in distributions to derive “subsampling scaling” that allows to extrapolate the results from subsampling to a full system: P(s) = psubPsub(s/
psub) where P(s) is the original distribution, Psub is the one under subN
sampling, and psub = M
is the probability to sample a unit, N—
number of sampled units, M—system size. In the model with critical
avalanches, subsampling scaling collapses distributions for any N
(Fig. 47B). However, for subcritical models, no distribution collapse is
observed (Fig. 47D). Thus we demonstrate that subsampling scaling
allows to distinguish critical from non-critical systems. With the help of
this novel method we studied dissociated cortical cultures. For these
we artificially subsampled recordings by considering only fraction of
all 60 electrodes. We find that in the first days subsampling scaling
does not collapse distributions well, whereas mature cultures (~from
day 21) allow for a good collapse, indicating development toward criticality (Fig. 47C, E).
Acknowledgements: AL received funding from the Marie Curie
Actions (FP7/2007–2013) under REA Grant Agreement No. (291734).
VP received funding from BMBF Bernstein 01GQ1005B.
References
1. Stumpf, MPH, Wiuf C, May RM. Subnets of scale-free networks are not
scale-free: sampling properties of networks. PNAS 2005;102(12):4221–24.
2. Priesemann V, Munk MH, Wibral M. Subsampling effects in neuronal
avalanche distributions recorded in vivo. BMC Neurosci. 2009;10(1):40.
3. Uhlig M, Levina A, Geisel T, Herrmann JM. Critical dynamics in associative
memory networks. Front Comp Neurosci. 2013;7.
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 47 Subsampling scaling in model and experiment. Left branching process model; right: experiments on developing cultures A
Avalanche size counts f(s) from the full and the subsampled critical
model; N: number of sampled neurons. B Under subsampling scaling,
all f(s) collapse. C Collapse of subsampled avalanche-size distribution
from the culture at the age of 21 days. D For subcritical models, the
same scaling ansatz does not result in a collapse. E No collapse of f(s)
from the culture at age 7 days
P77
Influences of embedding and estimation strategies on the
inferred memory of single spiking neurons
Lucas Rudelt1, Joseph T. Lizier2, Viola Priesemann1
1
Department of Non‑linear Dynamics, Max Planck Institute for Dynamics
and Self‑Organization, Göttingen, Germany; 2School of Civil Engineering,
The University of Sydney, Sydney, NSW, Australia
Correspondence: Lucas Rudelt ‑ l.rudelt@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P77
Information theory provides a generic framework for studying statistical dependencies, and is widely used in neuroscience.
However, a correct estimation of the involved quantities can be
challenging. For example, a correct estimation of transfer entropy
(TE), TE(X → Y) = I(X−, Y|Y−), or active information storage (AIS),
AIS(X) = I(X−, X), requires past state variables X− and Y− that encode
all information about the past that is relevant when predicting X or Y
[1]. For a spiking neuron, states can be defined by transforming the
spike train into a binary sequence of spike counts in sufficiently small,
equally spaced bins with some bin size Δt (Fig. 48A). For neurons, however, it is unclear how many past bins a sufficient state variable typically comprises. In practice, past states have often been limited to only
one time bin to reduce the complexity of estimation. This points at the
main challenge one faces when estimating TE and AIS: A reliable estimation of probabilities from recorded data becomes more and more
difficult with increasing complexity of the state variables, i.e. with
considering more past bins. We used AIS for single spiking neurons to
estimate (a) how much memory there is, (b) how long it reaches typically into the past, (c) the (non-)linear contributions of the memory. To
this end, we first examined the performance of different estimators
Page 51 of 112
Fig. 48 Relative active information storage as a function of the time
range of the past state for different estimators
for a realistic model neuron [2] whose AIS can be directly computed
(dashed line, Fig. 48B). Using constant external drive we simulated a
recording of 12 h. In a model-free approach, probabilities were directly
estimated from relative frequencies using the standard ‘plugin’ estimator or the ‘NSB’ estimator [3]. In addition, we fitted a generalized
linear model (GLM) whose predictions constitute an estimator that is
constrained to linear contributions. For all these estimation strategies,
the number of past bins k and thus the time range was systematically
varied (Fig. 48B). We then applied the same estimators to in vitro and
in vivo recordings (Fig. 48c, d) of 3 h and 1 h duration.
Considering a very small number k of past bins can lead to a substantial underestimation of AIS. Increasing the number of past bins,
however, leads to severe positive bias of the model-free estimators
when the complexity of the past state becomes too large. This manifests in the estimators exceeding the true AIS (Fig. 48B). Assuming
a point process with linear contributions (GLM), in contrast, allows
a robust estimation but does not capture non-linear effects. This
approach can also be used to take the global past activity of the
neural population into account, thereby unveiling redundancies in
the activity of the single neuron with the population activity. While
the model neuron intrinsically has only linear dependencies on its
past, it is surprising that, in vitro, there also seem to be very little
non-linear contributions and a lot of redundancy. In vivo, the nonlinear contributions are more prominent and the memory is clearly
non-redundant.
We thus showed that appropriate embedding is necessary, otherwise
AIS is underestimated and likewise, TE might be mis-estimated. Furthermore, our results suggest that in vivo the information processing
is more evolved.
References
1. Wibral M, Lizier J, Priesemann V. Bits from brains for biologically-inspired
computing. Comput Intell. 2015;2:5.
2. Pozzorini, C, Naud R, Mensi S, Gerstner W. Temporal whitening by powerlaw adaptation in neocortical neurons. Nat Neurosci. 2013;16(7):9428.
3. Nemenman I, Shafee F, Bialek W. Entropy and inference, revisited. Adv
Neural Inf Process Syst. 2002;14:471–478.
BMC Neurosci 2016, 17(Suppl 1):54
P78
A nearest‑neighbours based estimator for transfer entropy
between spike trains
Joseph T. Lizier1, Richard E. Spinney1, Mikail Rubinov2,3, Michael Wibral4,
Viola Priesemann5,6
1
Complex Systems Research Group, Faculty of Engineering & IT, The
University of Sydney, NSW 2006, Australia; 2Janelia Research Campus,
Howard Hughes Medical Institute, Ashburn, VA 20147, USA; 3Department
of Psychiatry, University of Cambridge, Cambridge, UK; 4MEG Unit,
Brain Imaging Center, Goethe University, 60528 Frankfurt am Main,
Germany; 5Department of Nonlinear Dynamics, Max Planck Institute
for Dynamics & Self‑Organization, Göttingen, Germany; 6Bernstein Center
for Computational Neuroscience, Göttingen, Germany
Correspondence: Joseph T. Lizier ‑ joseph.lizier@sydney.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P78
The nature of a directed relationship (or lack thereof ) between brain
areas is a fundamental topic of inquiry in computational neuroscience
[1]. A particular focus of such inquiry is in regards to the analysis of
information flows in a network [2], and such investigations take place
at all levels of analysis, from interregional connectivity in fMRI imaging data [3] down to directed relationships between spike trains at the
neuronal level [4].
In all of the aforementioned studies, information theory provides the
primary tool, transfer entropy (TE) [5], for analysis of such directed relationships. TE measures the predictive gain about state transitions in a
target time-series from observing some source time-series. While the
TE has been used extensively to analyse recordings from fMRI, MEG
and EEG for example [1–3], fewer applications [4] have been made to
spiking time-series. Although one can apply temporal binning on such
time-series before measuring TE on the resultant binary time-series
[4], it remains unclear: (a) how to set parameters for this approach (e.g.
bin sizes), (b) whether an estimate can be achieved by avoiding temporal binning and instead working directly with continuous-valued
time stamps of spikes, and (c) whether such an estimate would actually improve on binning approaches.
Recent theoretical developments have pointed to how transfer
entropy may be derived from continuous-valued time-stamps of
spikes directly, using spike rates conditioned on previous spike histories [6]. Yet, it is not immediately obvious how an estimator for this
form would be constructed, and indeed construction of such an estimator has previously defaulted to a binning or discretisation of time
[7]. Here, we propose an estimator for this continuous-time pointprocess formulation of TE that remains in the continuous-time regime
by harnessing a nearest-neighbours approach [8] to matching (rather
than binning) inter-spike interval (ISI) histories and future spike-times.
By retaining as much information about ISIs as possible, this estimator
is expected to improve on properties of TE such as robustness to noise
and undersampling, bias removal, and sensitivity, etc. We are currently
implementing the proposed estimation algorithm in open-source
code (i.e. contributing to JIDT [9] and TRENTOOL [10]), and evaluating
the properties of the algorithm particularly in comparison to temporal
binning approaches.
References
1. Wibral M, Vicente R, Lizier JT, editors. Directed information measures in
neuroscience. Berlin: Springer-Verlag; 2014.
2. Vicente R, Wibral M, Lindner M, Pipa G. Transfer entropy—a model-free
measure of effective connectivity for the neurosciences. J Comp Neurosci. 2011;30(1):45–67.
3. Lizier JT, Heinzle J, Horstmann A, Haynes J-D, Prokopenko M. Multivariate
information-theoretic measures reveal directed information structure
and task relevant changes in fMRI connectivity. J Comp Neurosci.
2011;30(1):85–107.
4. Ito S, Hansen ME, Heiland R, Lumsdaine A, Litke AM, Beggs JM. Extending
transfer entropy improves identification of effective connectivity in a
spiking cortical network model. PLoS One. 2011;6(11):e27431.
5. Schreiber T. Measuring information transfer. Phys Rev Lett. 2000;85:461–4.
6. Bossomaier T, Barnett L, Harré M, Lizier JT. An introduction to transfer
entropy: information flow in complex systems. Berlin: Springer; 2016 (in
press).
Page 52 of 112
7.
Kim S, Putrino D, Ghosh S, Brown EN. A Granger causality measure for
point process models of ensemble neural spiking activity. PLoS Comput
Biol. 2011;7(3):e1001110.
8. Kraskov A, Stögbauer H, Grassberger P. Estimating mutual information.
Phys Rev E. 2004;69(6):066138.
9. Lizier JT. JIDT: an information-theoretic toolkit for studying the dynamics
of complex systems. Front Robot AI. 2014;1:11.
10. Lindner M, Vicente R, Priesemann V, Wibral M. TRENTOOL: a Matlab open
source toolbox to analyse information flow in time series data with transfer entropy. BMC Neurosci. 2011;12(1):119.
P79
Active learning of psychometric functions with multinomial
logistic models
Ji Hyun Bak1, Jonathan Pillow2
1
Department of Physics & Lewis‑Sigler Institute for Integrative
Genomics, Princeton University, Princeton, NJ 08544, USA; 2Department
of Psychology & Princeton Neuroscience Institute, Princeton University,
Princeton, NJ 08544, USA
Correspondence: Ji Hyun Bak ‑ jhbak@princeton.edu
BMC Neuroscience 2016, 17(Suppl 1):P79
As new technologies expand the capacity for making large-scale
measurements of neural activity, there is a growing need for methods to rapidly characterize behavior and its dependence on stimuli. In
typical experiments, an animal is presented with a stimulus on each
trial and has to select a response among several options. Since such
experiments are costly, a problem of practical importance is to learn
the animal’s psychometric choice functions from a minimal amount of
data. Here we show that one can achieve substantial speedups over
traditional randomized designs via active learning, in which stimuli are
selected adaptively on each trial according to an information-theoretic
criterion, as shown in Fig. 49. Specifically, we model behavior with a
multinomial logistic regression model, in which the probability of
each choice given a stimulus depends on a set of linear weights. Our
work extends previous work on this problem [1–3] in several important ways. First, we incorporate an explicit lapse rate to account for the
fact that observers may occasionally make errors on “easy” trials due
to lapses in concentration or memory [4]. Second, we develop an efficient method based on Markov Chain Monte Carlo (MCMC) sampling
that is accurate in settings in which the log-likelihood is not concave,
for example as in the presence of lapse rates. Third, we extend consideration for multiple-alternative responses, extending previous work
for binary responses. We compare the performance of our samplingbased method to one based on a local (Laplace) approximation to
Fig. 49 Example of active learning, simulated with a three-alternatives model on 1D stimulus. After each observation, the psychometric
functions are estimated based on the accumulated data, and the next
stimulus is chosen to maximize the expected information gain. The
estimated psychometric functions (solid lines) quickly approach the
true functions (dashed lines) through the adaptive and optimal choice
of stimuli
BMC Neurosci 2016, 17(Suppl 1):54
the posterior [5], and show that failure to incorporate lapse rates can
have deleterious effects on the accuracy of inferred parameters under
both methods. We test our method on simulated data, as well as on
an experimental dataset concerning the multiple-alternative choice
behavior of monkeys [6], demonstrating that active sampling of the
stimulus space facilitates the learning of the psychometric function
significantly, as well as suggesting that the full range of the multidimensional stimulus could have been exploited more efficiently
using our active learning framework. Finally, we discuss the comparative advantages and disadvantages of the different methods, and how
one might adapt these algorithms to achieve best results.
Acknowledgements: We thank Anne Churchland for providing the
data. JP was supported by the McKnight Foundation, Simons Global
Brain Initiative, NSF CAREER Award IIS-1150186, and NIMH grant
MH099611. JHB was supported by NSF grant PHY-1521553.
References
1. Kontsevich LL, Tyler CW. Bayesian adaptive estimation of psychometric
slope and threshold. Vis Res. 1999;39:2729–37.
2. Zocchi SS, Atkinson AC. Optimum experimental designs for multinomial
logistic models. Biometrics. 1999;55:437–44.
3. DiMattina C. Fast adaptive estimation of multidimensional psychometric
functions. J Vis. 2015;15:1–20.
4. Kuss M, Jakel F, Wichmann FA. Bayesian inference for psychometric functions. J Vis. 2005;5:478–92.
5. Lewi J, Butera R, Paninski L. Sequential optimal design of neurophysiology
experiments. Neural Comput. 2009;21:619–87.
6. Churchland AK, Kiani R, Shadlen MN. Decision-making with multiple
alternatives. Nat Neurosci. 2008;11:693–702.
P81
Inferring low‑dimensional network dynamics with variational
latent Gaussian process
Yuan Zaho1,2, Il Memming Park1,3
1
Department of Neurobiology and Behavior, Stony Brook University,
Stony Brook, NY 11794, USA; 2Department of Applied Mathematics
and Statistics, Stony Brook University, Stony Brook, NY 11794, USA;
3
Institute for Advanced Computational Science, Stony Brook University,
Stony Brook, NY 11794, USA
Correspondence: Il Memming Park ‑ memming.park@stonybrook.edu
BMC Neuroscience 2016, 17(Suppl 1):P81
Surprisingly, large-scale population recordings often show signatures
of low-dimensional dynamics, that is, variations in a small number of
common factors explain most of the dependence among neurons
[1–3]. This supports the idea that a large neuronal network is implementing necessary computations described by continuous lowdimensional nonlinear dynamics. Sufficient amount of redundancy in
the population activity would allow us access to the internal computation process of interest even when we only observe a small subset of
neurons. Thus, it is necessary to deduce the latent dynamics from neural time series in order to understand if and how neural systems operate in this regime. There are several latent variable models that aim
at recovering the latent dynamics, howerver, they make inadequate
assumptions in favor of fast inference [4]. Here we describe an approximate inference method that recovers the latent dynamics under a natural generative model with minimal assumptions.
We implemented a probabilistic method to extract shared low-dimensional latent dynamics from multi-channel neural recordings (LFP and
spike trains) to reveal how neural population encodes information,
and how multiple functional neural populations dynamically interact
with each other. Key assumptions of our model are: (1) each neural
signal represent a noisy mixture of common latent dynamics, and (2)
latent dynamics are independent and temporally smooth (with possibly different time scales). We use autoregressive generalized linear
model driven by latent dynamics. Unlike most of the literature [5], we
do not impose linear dynamics as a prior on the latent process, instead
we use a general gaussian process prior which provides a flexible
framework for imposing structure such as smoothness. However, as a
Page 53 of 112
result, the exact posterior inference is intractable, thus we developed
a variational method to find a Gaussian approximation to the posterior
[6]. Our inference algorithm is memory-efficient and fast: both linear in
time using a low-rank approximation of the covariance. We compare
our method on both simulated systems and real data from V1 driven
by drifting gratings. For a population of 148 V1 neurons, 11.4 % of
the variance was explained by a shared 4-dimensional latent process,
while 10 % of the variance was explained by independent variability
of each neuron. We recovered orientation dependent embedding that
faithfully encode the stimulus drive on average, and the populationwide trial-to-trial modulation. In conclusion, we present an efficient
and scalable method to recover underlying dynamics from noisy partial observations to study neural code and computation.
References
1. Okun M, Steinmetz NA, Cossell L, Iacaruso MF, Ko H, Barthó P, Moore T,
Hofer SB, Mrsic-Flogel TD, Carandini M, Harris KD. Diverse coupling of
neurons to populations in sensory cortex. Nature. 2015;521(7553):511–15.
2. Goris RLT, Movshon JA, Simoncelli EP. Partitioning neuronal variability. Nat
Neurosci. 2014;17(6):858–65.
3. Luczak A, Bartho P, Harris KD. Gating of sensory input by spontaneous
cortical activity. J Neurosci. 2013;33(4):1684–95.
4. Yu BM, Cunningham JP, Santhanam G, Ryu SI, Shenoy KV, Sahani M.
Gaussian-process factor analysis for low-dimensional single-trial analysis
of neural population activity. J Neurophysiol. 2009;102(1):614–35.
5. Paninski L, Ahmadian Y, Ferreira DGG, Koyama,S, Rahnama Rad K, Vidne
M, Vogelstein J, Wu W. A new look at state-space models for neural data.
Journal of Comput Neurosci. 2010;29(1–2):107–26.
6. Blei DM, Kucukelbir A, McAuliffe JD. Variational inference: a review for statisticians. arXiv 2016 1601.00670 [http://arxiv.org/abs/1601.00670 arXiv]
P82
Computational investigation of energy landscapes in the resting
state subcortical brain network
Jiyoung Kang1, Hae‑Jeong Park2
1
Graduate School of Life Science, University of Hyogo, 3‑2‑1 Koto,
Kamigori, Ako, Hyogo, 678‑1297, Japan; 2Department of Nuclear
Medicine, Radiology and Psychiatry, Yonsei University College
of Medicine, Department of Cognitive Science, Yonsei University, 50
Yonsei‑ro, Sinchon‑dong Seodaemoon‑gu, Seoul, 120‑752, Republic
of Korea
Correspondence: Hae‑Jeong Park ‑ parkhj@yuhs.ac
BMC Neuroscience 2016, 17(Suppl 1):P82
Recently, energy landscapes of the resting state functional brain network have been researched using the pairwise maximum entropy
model (MEM). This approach considers not only activities of nodes but
also interactions among nodes in modeling brain networks and thus
estimating energy landscapes of the brain state [1, 2]. From the energy
landscape models, we can identify major stable states (local minima)
and estimate transition rates among stable states.
The brain networks of the resting states are known to be affected by
brain diseases or treatments. In the pairwise MEM, such effects correspond to changes in the parameters of the baseline activities and pairwise interactions.
In the present study, we investigated the energy landscape and its
robustness of the subcortical human brain network that plays a central
role in the human brain. The subcortical brain regions we examined
were 15 regions of interests (ROIs); hippocampus, amygdala, caudate,
putamen, pallidum, thalamus, nucleus accumbens, and brainstem. To
construct a pairwise MEM for spontaneous interactions among subcortical brain regions, we used resting state fMRI (rs-fMRI) data of the
human connectome projects, which contains 468 people’s data. The
blood oxygen level-dependent signals in the ROI were first binarized
to represent states (zero for inactive, one for active states) of the ROI,
and thereby 215 brain states were considered. The parameters of the
MEM were fitted to reproduce observed activation patterns of the rsfMRI data. The constructed MEM showed high accuracy of fit (~92.6 %)
and reliability (~99.9 %).
BMC Neurosci 2016, 17(Suppl 1):54
We found symmetric properties for the left and right hemispheres, and
confirmed estimated parameters grossly reflecting the known anatomical connectivity of the subcortical brain. We further investigated
the robustness of the system by perturbing the global weight for interactions, parameters for baseline activities of ROI and parameters for
interactions between pairs of all ROIs (1906 edges), one by one from
the original MEM. Alteration of the energy landscapes after perturbation was measured with respect to the number of local minima. We
found that the number of the local minima of the subcortical system
without any perturbation is very high. This implies that the subcortical brain system is optimal in the sense of its largest coverage of local
minima (maximal number of local minima). This result suggests that
brain was built to have multiple stable states. We also found different categories of parameters that affect the energy landscape of the
resting state. For example, small increase in the pairwise parameter
between the caudate and putamen dramatically reduced the numbers
of the local minima while reduction in this parameter did not change
the energy landscape.
In conclusion, MEM analysis of resting state functional network would
be an important tool to understand principles of the brain organization and could be useful in researching brain disease.
References
1. Watanabe T, Hirose S, Wada H, Imai Y, Machida T, Shirouzu I, Konishi S,
Miyashita Y, Masuda N. A pairwise maximum entropy model accurately describes resting-state human brain networks. Nat Commun.
2013;4:1370.
2. Watanabe T, Hirose S, Wada H, Imai Y, Machida T, Shirouzu I, Konishi S,
Miyashita Y, Masuda N. Energy landscapes of resting-state brain networks.
Front Neuroinform. 2014;8:12.
P83
Local repulsive interaction between retinal ganglion cells can
generate a consistent spatial periodicity of orientation map
Jaeson Jang1, Se‑Bum Paik1,2
1
Department of Bio and Brain engineering, Korea Advanced Institute
of Science and Technology, Daejeon 34141, Republic of Korea; 2Program
of Brain and Cognitive Engineering, Korea Advanced Institute of Science
and Technology, Daejeon 34141, Republic of Korea
Correspondence: Jaeson Jang ‑ jaesonjang@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P83
Orientation map in the primary visual cortex (V1) is of great interest
among functional maps in the brain, but its developmental mechanism has been under debate. A recently suggested idea is that a moiré
interference pattern between ON and OFF retinal ganglion cells (RGCs)
can develop a quasi-periodic structure of orientation map [1] (Fig. 50A).
In this model, the mosaics of ON and OFF RGCs that are in hexagonal
lattice patterns generate a periodic interference pattern and induce a
cortical orientation preference map. This model successfully explains
the mechanism of map development, but two questions remain unanswered yet; (1) How does the hexagonal pattern of RGC mosaic develop
and (2) how is the angle alignment (θ) between ON and OFF RGC
mosaics (Fig. 50A) restricted to seed the consistent spatial periodicity
of orientation map? Here, we suggest that a local repulsive interaction
between the nearby cells is enough to develop hexagonal RGC mosaics
and consistent alignment of ON and OFF mosaics.
To validate this idea of developmental process of cell mosaic, we
assumed a local repulsive force between the nearby cells as a function of distance between two cells (Fig. 50B), which induces a gradual
shift of cell position. In our model simulations, we confirmed that
this model could develop a hexagonal pattern in the monotypic RGC
mosaic (Fig. 50C, D). Next, we examined how the angle alignment
between ON and OFF mosaics can be achieved by homotypic (ON–ON
or OFF–OFF) and heterotypic (ON–OFF) interaction between RGCs. We
simulated the development of ON and OFF mosaics as we allow a heterotypic interaction and gradually reduce the distance between two
mosaics (Fig. 50E). When two mosaics get closer enough, we observed
that the angle alignment between ON and OFF mosaics was limited to
low angles (Fig. 50F). Finally, we analyzed previously reported cat RGC
mosaic data [2, 3] and concluded that the data suggests the existence
Page 54 of 112
Fig. 50 Local repulsive interaction develops a consistent interference
between mosaics. A Moiré pattern of RGC. B Developmental model
of RGC mosaic with local repulsive interaction between nearby cells.
C Developed cell mosaic. D Autocorrelation of developed mosaics. e
Approach between ON and OFF mosaics induces a gradual reinforcement of heterotypic interaction. F Angle alignment between mosaics
(θ) is limited to low angles as mosaics approach (*: p < 0.05, Ranksum
test, error bar: SE)
of heterotypic repulsive interaction between ON and OFF mosaics, as
our model predicted.
Our results suggest that a local repulsive interaction between RGCs can
develop a hexagonal pattern in mosaics and restrict the angle alignment between ON and OFF RGC mosaics to generate a constant spatial
period of orientation map. This model may provide a complementary
mechanism of the retinal origin of periodic functional maps in the brain.
References
1. Paik S-B, Ringach DL. Retinal origin of orientation maps in visual cortex.
Nat Neurosci. 2011;14:919–925.
2. Zhan XJ, Troy JB. Modeling cat retinal beta-cell arrays. Vis Neurosci.
2000;17:23–39.
3. Wassle H, Boycott BB, Illing R-B. Morphology and mosaic of ON- and OFFbeta cells in the cat retina and some functional considerations. Proc R Soc
B Biol Sci. 1981;212:177–95.
P84
Phase duration of bistable perception reveals intrinsic time scale
of perceptual decision under noisy condition
Woochul Choi1,2, Se‑Bum Paik1,2
1
Department of Bio and Brain Engineering, KAIST, Daejeon 34141,
Republic of Korea; 2Program of Brain and Cognitive Engineering, KAIST,
Daejeon 34141, Republic of Korea
Correspondence: Woochul Choi ‑ choiwc1128@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P84
When we see an ambiguous visual stimulus such as Necker cube, our
perceived state switches periodically between two possible interpretations. This phenomenon, called bistable perception, is considered
important to the study of dynamic mechanism of sensory perception.
In particular, the time duration of each perceptual state, termed “phase
duration”, seems to be a crucial factor to understanding the temporal
features of underlying neural activity during sensory perception under
this condition, which has not been studied intensively yet.
In this study, we assume that phase duration is intrinsically correlated with time delay in cognitive tasks, such as decision making from
sensory information. Our hypothesis is that the periodic switching
in bistable perception is a repeated process of decision making and
reveals the time scale required for this decision task. To confirm our
hypothesis, we performed a human psychophysics experiment using
the “racetrack” type stimulus [1], which can induce a motion perception from both bistable illusion and real motion by varying the coherence parameter, c (Fig. 51A). We examined the relationship between
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 51 Correlation between bistable perception and perception
under ambiguous signal. A Racetrack stimulus. Rotational motion can
be either illusory or ambiguous depending on coherence. B Example
response of racetrack. Perceived motion can be bistable (top) or
follows actual motion with response time (bottom). C Subjects’
(black) and model’s (red) phase duration and response time are highly
correlated. D Double-well energy model to describe behavior during
bistable perception and perceptual decision making task
the phase duration, τ, under illusory motion (c = 0) and the response
delay for coherent motion with different degrees of ambiguity (c > 0,
Fig. 51B). Our result showed that the response delay in the coherent
motion detection task (c > 0) was highly correlated with the phase
duration in the bistable illusory motion perception (c = 0) (Fig. 51C,
N = 19, R = 0.61, p < 0.01, Pearson’s correlation coefficient). For a
systematic analysis of subjects’ performance for these two tasks, we
designed a theoretical model of simple double-well energy potential
[2] (Fig. 51D). The model could successfully replicate the correlation
between phase duration and response delay in each task, suggesting
that bistable perception and perceptual decision making processes
may share a common neural mechanism (Fig. 51C).
Conclusions Our findings show that the phase duration of bistable
perception is highly correlated with the response time of a cognitive
task. Our simple model suggests that the bistable perception can
be interpreted as perceptual decision making process under highly
ambiguous condition and share similar temporal dynamics.
Page 55 of 112
monkeys and cats, preferred orientation of each V1 neuron appears
continuous and periodic across cortical space. On the other hand, in
rodents, it appears completely discontinuous, forming a structure
called salt-and-pepper map. However, the developmental mechanism
of salt-and-pepper orientation maps remains unclear. Previously, a
model study suggested that a moiré interference pattern between ON
and OFF retinal ganglion cell (RGC) mosaics can seed a periodic orientation maps (Fig. 52A), and a salt-and-pepper map can also be developed when the spatial periodicity of interference pattern is very short
[1]. However, our analysis suggests that the spatial periodicity of map,
estimated from rat RGC mosaics data, is not small enough to generate
a salt-and-pepper structure (Fig. 52B, C). To address this issue, here we
suggest that feedforward convergent wiring between retina and V1 is
a crucial factor that decides the structure of orientation map.
To find a convergence condition that develops salt-and-pepper map,
we modulated two parameters in our simulations: (1) the convergence
range of V1 cells and (2) the sampling ratio of RGCs within the range.
The regularity of orientation map was estimated from the measurement of the preferred orientation difference between local neurons
(Fig. 52D). We found that a salt-and-pepper map was developed with
low sampling ratio and large convergence range, while a smooth map
was developed when convergence range was relatively small and
sampling ratio was high (Fig. 52E). To further analyze the map structure generated by our model, we compared the profile of correlation
between local receptive fields structure in our simulated V1 map to the
previous observation in animal experiment [2]. We confirmed that our
salt-and-pepper map model well matched the statistics of observed
experimental data.
Conclusions Our result suggests that a salt-and-pepper map can be
developed by sparse and long-range convergence in feedforward wiring, while smooth map can be developed by localized convergence.
We suggest that the condition of feedforward convergence between
retina and V1 is a critical factor to determine the structure of orientation map.
References
1. Jain S. Performance characterization of Watson Ahumada motion detector using random dot rotary motion stimuli. PLoS One. 2009;4.
2. Kornmeier J, Bach M. Ambiguous figures—what happens in the brain
when perception changes but not the stimulus. Front Hum Neurosci.
2012;6:51.
P85
Feedforward convergence between retina and primary visual
cortex can determine the structure of orientation map
Changju Lee1, Jaeson Jang1, Se‑Bum Paik1,2
1
Department of Bio and Brain Engineering, Korea Advanced Institute
of Science and Technology, Daejeon 34141, Republic of Korea; 2Program
of Brain and Cognitive Engineering, Korea Advanced Institute of Science
and Technology, Daejeon 34141, Republic of Korea
Correspondence: Changju Lee ‑ lcj110808@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P85
Orientation map in the primary visual cortex (V1) is one of the most
studied functional maps in the brain. In higher mammals such as
Fig. 52 The simulation model for developmental mechanism of saltand-pepper map by feedforward convergence between retina and
V1. A Moiré interference between ON and OFF RGC mosaics. B Moiré
interference with small and large alignment angles can generate
various range of periodicity, S. C From rat RGC mosaics, both smooth
and salt-and-pepper map can be developed. D Spatial distribution of
preferred orientations of V1 cells by different convergence conditions;
Smooth map model (high sampling ratio and large convergence
range), Salt-and-pepper map model (low sampling ratio and short
convergence range). E The structure of orientation map depending
on convergence range and sampling ratio
BMC Neurosci 2016, 17(Suppl 1):54
P86
Computational method classifying neural network activity
patterns for imaging data
Min Song1,2, Hyeonsu Lee1, Se‑Bum Paik1,2
1
Department of Bio and Brain Engineering, KAIST, Daejeon 34141,
Republic of Korea; 2Program of Brain and Cognitive Engineering, KAIST,
Daejeon 34141, Republic of Korea
Correspondence: Min Song ‑ night@kaist.ac.kr, Hyeonsu Lee ‑
hslee9305@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P86
In neural imaging data, various types of spatio-temporal activity patterns are observed which may reflect dynamic features of information processing in the brain [1, 2]. Classification of these patterns is
required to analyze the brain activity further. However, development
of analysis tool for spatio-temporal neural activity pattern has been
regarded difficult because of highly complex connections between
neurons and nonlinearity of activity patterns. In this study, we suggest a novel method of classifying activity patterns in two aspects:
spatial geometry and temporal dynamics. We show that our method
efficiently categorizes complicated spatio-temporal patterns in brain.
First, we defined meaningful activity as salient distribution of highly
activated parts. Its spatial feature could be described by size and
peak amplitude (Fig. 53A), temporal feature by velocity and dispersion of activity (Fig. 53A). Thus, we designed geometric profile as twodimensional profile of instantaneous neural activity, by measuring
the topography of supra-threshold area with a shifting threshold. This
profile contains the information of size, peak amplitude, and geometric contours of meaningful activity. With this method, we could readily
estimate similarity or correlation of different activities in terms of size,
peak amplitude, and amplitude contour (Fig. 53B).
Next, we defined propagation profile as a characteristic of temporal displacement of activity on each direction against time and angle
axis. We measured trajectory and speed of activity using a normalized
cross-correlation. This profile intuitively shows dominant trajectory,
speed change and dispersion of the activity: how disperse to every
direction. So we can compare activities: whether the activity is moving
straight or curved trajectories, or whether the activity propagation is
accelerating or not (Fig. 53C).
Our new method can easily perform not only classification of overall dynamics in brain, but also a simplified description of complex
patterns (Fig. 53D), that may be applicable to the analysis of various
kinds of brain imaging data.
P87
Symmetry of spike‑timing‑dependent‑plasticity kernels regulates
volatility of memory
Youngjin Park1, Woochul Choi1,2, Se‑Bum Paik1,2
1
Department of Bio and Brain Engineering, Korea Advanced Institute
of Science and Technology, Daejeon 34141, Republic of Korea; 2Program
of Brain and Cognitive Engineering, Korea Advanced Institute of Science
and Technology, Daejeon 34141, Republic of Korea
Correspondence: Youngjin Park ‑ yodamaster@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P87
Synaptic plasticity is considered the core mechanism of learning and
memory [1]. However, how plasticity can specifically modulate synaptic connections to generate short term or long term memory has not
been understood completely. Here we introduce a theoretical model
which suggests that a key mechanism of short term and long term
memory can be implemented by a small difference in spike-timingdependent-plasticity (STDP) rule. (Fig. 54A).
To test our idea, we designed simulations using a model feedforward
neural network where two types of synaptic plasticity are implemented; asymmetric STDP (AS) [2] and symmetric STDP (SS). We
defined the memory as the ability of a system to retrieve a consistent
response spike pattern when we repeatedly introduce an identical
pattern of spikes, and then we examined the performance of the system in terms of memory sustainability and appendability.
In our simulations, a network with AS showed performances similar to
short-term memory while a network with SS showed long-term memory like properties. Memory in an AS Network decayed as a function of
time, while memory in a SS network did not show a noticeable decay
(Fig. 54B). Moreover, when a new input pattern was given to the network in addition to old memory, AS system replaced old memory with
new memory pattern (Fig. 54C), while SS system maintained the old
memory together with a newly trained memory (Fig. 54D).
Based on our findings, we suggest a new memory system called hybrid
memory that is capable of showing intermediate properties between
a long-term memory and a short-term memory (Fig. 54B, hybrid). This
model suggests that transition between short term and long term
memory might not be discrete but gradual.
Conclusions We have shown that our model network can implement
different types of memory performance from the variation of plasticity,
A
1
B
LTP
Δw
pre
post
LTD
pre
post
0
-50
C
Asymmetric STDP
D
Memory index
Memory index
0
50
1
0
Fig. 53 A novel index effectively describes different neural activity
patterns obtained from imaging. A Neural activity obtained from
optical imaging could be analyzed with appearance and propagation.
B Appearance index of four distinct sample. C Propagation index of
straight trajectory (top) and curved trajectory (bottom). D Propagation
Index of non-dispersive sample (top) and dispersive sample (bottom)
0
Δt (ms)
0
1
200 400 600 800 1000 1200 1400
Time (s)
Asymmetric
Symmetric
Hybrid
Memory index
References
1. Paik S-B, Ringach DL. Retinal origin of orientation maps in visual cortex.
Nat Neurosci. 2011;14:919–25.
2. Bonin V, Histed MH, Yurgenson S, Reid RC. Local diversity and fine-scale
organization of receptive fields in mouse visual cortex. J Neurosci.
2011;31:18506–21.
Page 56 of 112
0
1000
500
Decaying time (s)
Symmetric STDP
1
0
0
200 400 600 800 1000 1200 1400
Time (s)
p1
p2
p3
p4
p5
p6
p7
Fig. 54 Different learning rules reproduce volatile/nonvolatile
memory system. A Spike timing dependent plasticity. B Memory
decaying properties of different learning rule. Poisson spikes are given
for 1000 s to simulate decaying environment. C, D Multiple patterns
was given to the system every 200 s. C Memory performance of each
pattern in AS memory system. D Memory performance of each pattern in SS memory system
BMC Neurosci 2016, 17(Suppl 1):54
Page 57 of 112
or learning rule. Our results imply that the various types of memory
may be originated from a small difference in the shape of STDP kernel.
References
1. Bliss TV, Collingridge GL. A synaptic model of memory: long-term potentiation in the hippocampus. Nature. 1993;361(6407): 31–9.
2. Gütig R, Aharonov R, Rotter S, Sompolinsky H. Learning input correlations
through nonlinear temporally asymmetric Hebbian plasticity. J Neurosci.
2003;23(9):3697–3714.
P88
Effects of time‑periodic coupling strength on the first‑spike
latency dynamics of a scale‑free network of stochastic Hodgkin–
Huxley neurons
Ergin Yilmaz1, Veli Baysal1, Mahmut Ozer2
1
Department of Biomedical Engineering, Bülent Ecevit University,
Zonguldak 67100, Turkey; 2Department of Electrical and Electronics
Engineering, Bülent Ecevit University, Zonguldak 67100, Turkey
Correspondence: Ergin Yilmaz ‑ erginyilmaz@yahoo.com
BMC Neuroscience 2016, 17(Suppl 1):P88
It is still not known clearly which encoding mechanism that neurons
utilize for coding of sensory information. One of the proposed encoding mechanism is latency coding which suppose that first-spike latency
conveys much of the information about the stimulus. In this context,
Pankratova et al. [1] studied the effects of noise on the first-spike latency
dynamics of stochastic the Hodgkin–Huxley (HH) neuron, and obtained a
bell-shaped dependence of mean response time of the neuron on noise
intensity, emerging a phenomenon called “noise delayed decay” (NDD).
Later, this finding have been extensively studied by using complex neuronal networks [2]. On the other hand, neurons exchange information via
coupling at the special location called synapse. Thus, coupled neurons
in networks play a decisive role on the phenomenon occurring in neuronal networks. Majority of the studies examining the NDD effect assume
that coupling strength among coupled neurons is constant and the fact
that synapses are plastic, that is, coupling strength among neurons can
change with time, have been neglected. To present the effects of plasticity or time-varying-coupling strength Birzu et al. [3] studied the effects
of time-periodic coupling strength (TPCS) on the firing dynamics of a
globally coupled array of FHN neurons. Here, our aim is to present the
effects of the frequency of TPCS on the NDD phenomenon in a scale-free
network of HH neurons. We construct the network with N = 200 neurons modeled by a stochastic HH equation including ion channel noise,
and average degree of kavg = 4. We consider that the coupling strength
among coupled units changes with time-periodic fashion as proposed in
[3]. To measure the mean latency and jitter of the network, the first-spike
times of each neuron are recorded. For the comparison purpose, we give
the constant coupling strength effect on the latency dynamics of scalefree network. Obtained result are depicted in Fig. 55.
Conclusions It is seen that mean latency and jitter of the first-spike
times exhibit a damped sine wave dependence on the frequency of
TPCS, indicating that TPCS can significantly increase or decrease the
latency time which passes until sensing of the suprathreshold stimulus by each neuron at fixed intensity of channel noise (S = 100 μm2).
The frequencies of TPCS greater than ω = 2 m s−1 are not significantly
affect the mean latency and jitter of the network, as compared to the
constant coupling strength. As a result, with finely tuned values of the
frequency of TPCS input signal detection performance of the scalefree network can be prominently increased by mitigating the response
time of the each neuron.
References
1. Pankratova EV, Polovinkin AV, Mosekilde E. Resonant activation in a stochastic Hodgkin–Huxley model: interplay between noise and suprathreshold driving effects. Eur Phys J B. 2005;45:391–7.
2. Yilmaz E. Impacts of hybrid synapses on the noise-delayed decay in
scale-free neural networks. Chaos Solitons Fractals. 2014;66:1–8.
3. Birzu A, Krischer K. Resonance tongues in a system of globally coupled
FitzHugh-Nagumo oscillators with time-periodic coupling strength.
Chaos. 2010;20(4):043114.
Fig. 55 The statistics of the first-spike occurrence times (amplitude
of TPCS ε0 = 0.2, cell size S = 100 μm2, frequency of suprathreshold
signal f = 20 Hz and amplitude of it A = 4μA/cm2). A Mean latency of
the network, B jitter of the network
P89
Spectral properties of spiking responses in V1
and V4 change within the trial and are highly relevant
for behavioral performance
Veronika Koren1,2, Klaus Obermayer1,2
1
Institute of Software Engineering and Theoretical Computer Science,
Technische Universitaet Berlin, Berlin, 10587, Germany; 2Bernstein Center
for Computational Neuroscience Berlin, Humboldt‑Universitaet zu Berlin,
Berlin, 10115, Germany
Correspondence: Veronika Koren ‑ veronika.koren@bccn‑berlin.de
BMC Neuroscience 2016, 17(Suppl 1):P89
Linking sensory coding and behavior is a fundamental question in
neuroscience. We have addressed this issue in behaving monkey visual cortex (areas V1 and V4) while animals were trained to perform a
visual discrimination task in which two successive images (target and
test stimuli, with a delay period in between) were either rotated with
respect to each other or were the same. We hypothesized that animal’s
performance in the visual discrimination task depends on the quality of stimulus coding in visual cortex. We tested this hypothesis by
investigating the power spectral density of spiking signal from single
neurons (spectra) and of pairs of neurons (cross-spectra) in relation
to correct and incorrect behavioral responses. Our analysis shows
that spectral properties systematically change with behavioral performance. Correct responses are associated with significantly higher
spectra during the delay period. Cross-spectra of correct responses
are significantly lower during the target period but significantly higher
afterwards (delay period and test period). Spectral properties of single
neurons and even more of pair-wise interactions therefore change
within the trial, presumably following functional demands of stimulus processing in different epochs of the trial. Interestingly, differential dynamics in visual cortex sustains successful versus unsuccessful
behavioral performance.
Preprocessing methods are used in order to avoid biases due to limited measurement time. The spike train is multiplied with Hanning
window for low frequencies up to 22 Hz and with Slepian multitapers
for frequencies between 24 and 140 Hz [1]. We use 300 ms window
of sustained activity during stimulus periods and 500 ms window of
activity during the delay period. Spectrum and cross spectrum are
computed with Fast Fourier transform (Matlab, Mathworks). The crossspectrum being a complex function, we consider its absolute value.
Spectra are averaged in bins of 6 Hz. The variance and the covariance
of the spiking signal are computed as sums over frequencies of auto
and cross-spectra, respectively, up to the cut-off frequency (140 Hz).
BMC Neurosci 2016, 17(Suppl 1):54
Page 58 of 112
Table 1 Estimated parameters for one-spike burster STG cell
fPar
C
gNa
ENa
gKd
Min
0.1
0
0
0
Max
10
800
100
200
Real
0.628
50
50
100
Estim.
0.613
62.07
39.05
95.77
EKd
gA
−100
0
0
75
−80
5
−76.09
4.622
EA
gCa
Ca0
Caf
−100
0
0.01
14
20
0
0
5
0.1
16
250
300
−80
4
0.05
14.96
200
250
0.0438
15.59
194.7
246.2
−88.59 3.912
Auto spectra are significantly higher in correct compared to incorrect
trials in most of frequency bands in both V1 and V4 areas (p < 0.05 in
21 and 22 out of 24 frequency bands in V1 and V4, respectively, onetailed sign-rank test). Consistently, the variance is significantly higher
for correct responses (p < 10−4 in V1 and p = 0.0014 in V4. Cross-spectra are lower in correct trials during target period (18 and 10 out of 24
frequency bands are significant in V1 and V4, respectively, no significant effect in remaining bands) but higher in delay period (11 and 20
bands are significant in V1 and V4, respectively, no effect in remaining bands) and test period (19 and 13 significant bands in V1 and V4,
respectively, no effect in remaining bands). Consistently, the covariance is significantly lower for correct responses during target stimulus (p = 0.0003, p = 0.0002 in V1 and V4) and higher during the delay
(p < 10−4 in V1 and V4) and the test stimulus (p < 10−4 in V1 and V4).
Our results show that spectra and cross spectra change during behavioral task and that spectral information in visual cortex might be highly
relevant for behavioral performance.
References
1. Fries P, Womelsdorf T, Oostenwald R, Desimone R. The effects of visual
stimulation and selective visual attention on rhythmic neuronal synchronization in Macaque area V4. J Neurosci. 2008;28(18):4823–35.
2. Womelsdorf T, Schoffelen JM, Oostenveld R, Singer W, Desimone R,
Engel AK, Fries P. Modulation of neuronal interactions through neuronal
synchronization. Science. 2007;316:1609–12.
3. Appel W. Mathematics for physics & physicists. Oxfordshire: Princeton
University Press; 2007.
4. Gutnisky DA, Dragoi V. Adaptive coding of visual information in neural
populations. Nature. 2008;452(7184):220–4.
P90
Methods for building accurate models of individual neurons
Daniel Saska1, Thomas Nowotny1
1
School of Engineering and Informatics, Sussex Neuroscience, University
of Sussex, Falmer, Brighton BN1 9QJ, UK
Correspondence: Daniel Saska ‑ research@saska.io
BMC Neuroscience 2016, 17(Suppl 1):P90
Formulating predictive models of single neuron dynamics has become
a challenge taken up by many researchers ever since Hodgkin and
Huxley published their widely accepted phenomenological model of
electrophysiological dynamics of the squid giant axon [1]. Advances
include, amongst others, modelling complex cells (such as cells of the
stomatogastric ganglion (STG) in lobster and crab [2]) or increasingly
automated modelling methods [3]. However, for each problem solved,
new ones emerge. One such problem has been pointed out by Golowasch et al. [4] who discovered that averaging multiple measurements
from the same cell type can produce models that fail to reproduce the
behaviour of the target cells. This issue does not only affect methods
that rely on averaging to achieve better signal-to-noise ratio but more
generally all methods that examine ion channels in separate preparations. This includes classical voltage clamp, in which different ionic
conductances are measured in separate individual cells because many
pharmacological blockers cannot be fully reversed.
We here propose a different approach for parameter estimation aiming to build a model based on data from a single, individual cell. The
proposed method consists of the consecutive use of a voltage clamp
like protocol and parameter estimation in a current clamp mode. For
Cat
gKCa
EKCa
−100
0
−80
gleak
Eleak
0
−100
1
0.01
−80.22 0.0106
0
−50
−50.15
the ‘voltage clamp protocol’ we use genetic algorithms (GA) to evolve
a set of ‘highlighting’ voltage waveforms and specific observation windows, so that the resulting currents within the windows depend on a
highlighted parameter but not so much on the values of other parameters. These parameter-specific waveforms are then applied to a live
neuron (so far in simulation) and the resulting currents are observed
and fitted with another GA, focusing on the highlighted parameters
for each of the voltage waveforms. The resulting model is then transferred into current clamp, where the parameters are again estimated
using a GA. The neuron models in the GA population are coupled
to the observed cell to achieve a degree of synchronization and so
smooth the error landscape. The coupling is reduced adiabatically
until the model neurons and experimental cell remain synchronized
with (virtually) no coupling.
We found that combining voltage and current clamp works particularly well since the fitness landscape in voltage clamp has few local
minima but is fairly shallow whereas the opposite is true for the current clamp mode. We can hence find approximate parameter values
from arbitrary initial guesses in our ‘voltage clamp’ mode and once the
parameters are in the right area, they can be refined in current clamp.
Our method can produce accurate models of cells in the crab STG for
the cases of one, two, three and four-spike bursters. Table 1 shows the
resulting parameter values for the example of a one-spike burster cell
illustrated in Fig. 56.
References
1. Hodgkin AL, Huxley AF. A quantitative description of membrane current
and its application to conduction and excitation in nerve. J Physiol.
1952;117(4):500–44.
2. Liu Z, Golowasch J, Marder E, Abbott LF. A model neuron with activitydependent conductances regulated by multiple calcium sensors. J
Neurosci. 1998;18(7):2309–20.
3. Willms AR. NEUROFIT: software for fitting Hodgkin–Huxley models to
voltage-clamp data. J Neurosci Methods. 2002;121(2):139–50.
4. Golowasch J, Goldman MS, Abbott LF, Marder E. Failure of averaging in
the construction of a conductance-based neuron model. J Neurophysiol.
2002;87(2):1129–31.
Fig. 56 One-spike burster and estimated model as in Table 1
BMC Neurosci 2016, 17(Suppl 1):54
Page 59 of 112
P91
A full size mathematical model of the early olfactory system
of honeybees
Ho Ka Chan1, Alan Diamond1, Thomas Nowotny1
1
School of Engineering and Informatics, University of Sussex, Falmer,
Brighton, BN1 9QJ, UK
Correspondence: Ho Ka Chan ‑ hc338@sussex.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P91
Experimental measurements often can only provide limited data from
animals’ sensory systems. As a result, data driven models are similarly
limited. However, in order to make biologically relevant predictions, it
is important to consider inputs representative of the full sensory input
space. Here we present a full size model of the early olfactory system
of honeybees that extrapolates inputs from the limited subset of available experimental observations.
Our model comprise olfactory receptor neurons (ORNs), local neurons (LNs) and projection neurons (PNs) organized in 160 glomeruli.
The ORN response patterns are generated using a set of ordinary differential equations describing the binding and activation of receptors
as in [2]. The parameters for these processes are chosen to match the
statistical distribution of experimental observed quantities in [3, 4] as
well as the statistics of asymptotic responses to time-invariant odours
at high concentration observed in calcium imaging of glomeruli with
bath-applied Ca dyes [5]. To generate the PN responses, we considered
a network in which PNs and LNs both receive excitatory input from
ORNs in the same glomerulus and inhibitory input from LNs in all glomeruli. The connectivity between PNs and LNs is based on the correlation between the activities of their respective glomeruli as in [5]. A
rate model, derived using the leaky integrate-and-fire model with the
assumption of constant input, is used to determine the input–output
relationship.
We tested our ORN model with continuous stimuli and short pulses.
The average normalized ORN responses to a chemical stimulus are
qualitatively similar to that of biological ORNs measured by electroantennogram recordings [4] as shown in Fig. 57, except that the time
scale of response latency is a little smaller in the model. This can be
explained by the lack of temporal filtering of input conductance to
output spiking in our rate model. The responses of PNs driven by ORN
activity can be compared to calcium imaging data with back-filled
PNs [6], which confirms that the responses of our model PNs can replicate key features of those of biological PNs. With appropriate data,
our model can be generalized to the early olfactory systems of other
insects. It hence provides a possible basis for future numerical studies
of olfactory processing in insects.
Acknowledgements: This work was
RGP0053/2015 and EPSRC, EP/J019690/1.
supported
by
HFSP,
Fig. 57 Experimental ORN responses to stimulus measured by electro-antennogram recordings in [5] (top, black line) is qualitative similar
to the average normalized ORN responses to stimulus (1-hexanol at
concentration 0.1 M) predicted by our model (bottom)
References
1. Galizia CG, Sachse S, Rappert A, Menzel R. The glomerular code for odor
representation is species specific in the honeybee Apis mellifera. Nat
Neurosci. 1999;2(5):473–8.
2. Grémiaux A, Nowotny T, Martinez D, Lucas P, Rospars J-P. Modelling the
signal delivered by a population of first-order neurons in a moth olfactory
system. Brain Res. 2012;1434:123–35.
3. Rospars J-P, Lansky P, Chaput M, Duchamp-Viret P. Competitive and noncompetitive odorant interactions in the early neural coding of odorant
mixtures. J Neurosci. 2008;28(10):2659–66.
4. Szyszka P, Gerkin RC, Galizia CG, Smith BH. High-speed odor transduction
and pulse tracking by insect olfactory receptor neurons. Proc Natl Acad
Sci USA. 2014;111(47):16925–30.
5. Linster C, Sachse S, Galizia CG. Computational modeling suggests that
response properties rather than spatial position determine connectivity
between olfactory glomeruli. J Neurophysiol. 2005;93(6):3410–17.
6. Ditzen M. Odor concentration and identity coding in the antennal lobe
of the honeybee Apis mellifera (PhD thesis). Berlin: Freie Universität Berlin;
2005.
P92
Stimulation‑induced tuning of ongoing oscillations in spiking
neural networks
Christoph S. Herrmann1, Micah M. Murray2, Silvio Ionta2, Axel Hutt3,
Jérémie Lefebvre4
1
Research Center Neurosensory Science, Carl‑von‑Ossietzky University
Oldenburg, Oldenburg, Germany; 2The Laboratory for Investigative
Neurophysiology (The LINE), Department of Clinical Neurosciences
and Department of Radiology, University Hospital Center and University
of Lausanne, Lausanne 1011, Switzerland; 3Deutscher Wetterdienst, 63067
Offenbach, Germany; 4 Krembil Research Institute, University Health
Network, Toronto, Ontario M5T 2S8, Canada
Correspondence: Jérémie Lefebvre ‑ jeremie.lefebvre@uhnresearch.com
BMC Neuroscience 2016, 17(Suppl 1):P92
Rhythmic neural activity is believed to play a central role in neural
computation. Oscillatory brain activity has been associated with myriad functions such as homeostasis, attention, and cognition as well as
neurological and psychiatric disorders, including Parkinson’s disease,
schizophrenia, and depression [1]. Numerous studies have shown that
that non-invasive stimulation, such as repetitive transcranial magnetic
stimulation (rTMS) and Transcranial Alternating Direct Current Stimulation (TACS), provide the means of modulating large-scale oscillatory
brain dynamics by perturbing and/or entraining both resting state
and task activity [2]. These stimulation-induced perturbations of neural oscillations have been shown to alter cognitive performance and
perception, effects that are further known to depend on brain state
prior and during stimulation [3]. Yet, the surge of interest in these
approaches is compromised by the existence of complex interference
patterns between exogenous and endogenous dynamics.
To better understand oscillatory responses evoked during rhythmic
stimulation, we simulated a spiking cortical network built of excitatory and inhibitory cells, expressing resting state alpha synchrony and
subjected to pulsatile forcing at frequencies in the range of 1–100 Hz.
Varying stimulation parameters—such as frequency and amplitude—
we evaluated the influence of stimulation on the spectral properties of
the network’ global neuroelectric output. The network was composed
of recurrently connected Poisson neurons with propagation delays,
linear adaptation, spatially profiled and sparse synaptic connections
and noisy inputs. To model exogenous influences, we used continuous trains of phasic pulses and stimulated the network globally (all
neurons identically), to mimic TMS-like signals. For every stimulation
condition, we also measured the neurons mean firing rate, the mean
network spike coherence and non-linearity metric. Multiple spectral
patterns could be observed in the network’s responses, both in the
power and frequency domains, indicating a plurality of responses to
shifts in stimulation frequency and/or amplitude. Network responses
to slower/weaker stimulation were expectedly found to be shaped
by entrainment and resonance: resonance curves defining the amplitude of the system’s responses were revealed, alongside the characteristic Arnold tongues, where stimulus-locking can be achieved. The
BMC Neurosci 2016, 17(Suppl 1):54
Page 60 of 112
individual firing rates of the neurons and resulting spike coherence
(assessing the degree of spiking synchronization) were both strongly
tied to the stimulation forcing. In contrast, for stimulation frequencies
higher than 50 Hz, a different mechanism was found to dominate the
network dynamics: stimulation pulses shaped the system’s response
via a non-linear acceleration (NLA) on ongoing oscillatory activity. The
network peak frequency was gradually shifted, leading to a transition
from the alpha to the beta band, and for forcing parameters that did
not recruit neither resonance nor entrainment. Also, NLA led the network in a state of weak oscillatory power, where individual neurons
were found in a state of intense, irregular spiking. By investigating
closely the network non-linear interactions for each stimulation conditions, we found that high-frequency forcing induces synergetic and
non-linear, large-scale effects [4]. Our results provide new computational perspectives about the response of synchronous spiking neural
networks in which firing rates, spike coherence and emergent oscillatory activity can be exogenously modulated using dynamic inputs.
Taken together, our results suggest that the action of forcing on oscillating neural systems must be regarded as strongly non-linear, and
input features must be considered as control parameters.
To overcome this issue, we take one step towards realistic in silico
experimental settings by using structured virtual environments to
obtain rich sensory input to drive model neural systems using the
ROS-MUSIC toolchain [5]. It allows us to simulate robotic agents in
virtual 3D environments performing a realistic perceptual decision
task, which can be directly equated to experimental data. The robotic
simulation generates realistic and structured sensory data which is
encoded to spiking neural activity using a nonlinear encoding process,
as formalized in [6]. The encoded sensory data is then used as input to
a balance recurrent neural circuit.
In this study, we investigate the emergent dynamical features of neural activity when the agent is navigating a virtual T-maze. We observe
decision-specific sequences of neural activity akin to experimental
evidence [1], revealing possible processing strategies employed by
the neural substrate. Furthermore, we investigate the role of different
adaptation/plasticity mechanisms in shaping the system’s dynamics. In
order to equate our results with those of other studies, we attempt to
partition the network state-space into discrete activity clusters, which
carry relevant information that could potentially be used to drive reinforcement learning algorithms.
References
1. Wang X-J. Neurophysiological and computational principles of cortical
rhythms in cognition. Physiol Rev. 2010;90:1195–1268.
2. Thut G, et al. The functional importance of rhythmic activity in the brain.
Curr Biol. 2012;22:658–63.
3. Neuling T, et al. Orchestrating neuronal networks: sustained after-effects
of transcranial alternating current stimulation depend upon brain states.
Front Hum Neurosci. 2013;7:161.
4. Lefebvre J, et al. Stimulus statistics shape oscillations in non-linear recurrent neural networks. J Neurosci. 2015;35(7):2895–903.
Acknowledgements: We acknowledge partial support the Helmholtz
Alliance through the Initiative and Networking Fund of the Helmholtz
Association and the Helmholtz Portfolio theme “Supercomputing and
Modeling for the Human Brain”, EuroSPIN and the German Federal
Ministry for Education and Research (BMBF Grant 01GQ1343).
P93
Decision‑specific sequences of neural activity in balanced random
networks driven by structured sensory input
Philipp Weidel1, Renato Duarte1,4,5, Abigail Morrison1,2,3,4
1
Institute of Advanced Simulation (IAS‑6) & Institute of Neuroscience
and Medicine (INM‑6) & JARA BRAIN Institute I, Jülich Research Center,
52425 Jülich, Germany; 2Institute of Cognitive Neuroscience, Faculty
of Psychology, Ruhr‑University Bochum, 44801 Bochum, Germany;
3
Simulation Laboratory Neuroscience – Bernstein Facility for Simulation
and Database Technology, Institute for Advanced Simulation, Jülich
Aachen Research Alliance, Jülich Research Center, Jülich, Germany;
4
Bernstein Center Freiburg, Albert‑Ludwig University of Freiburg,
Freiburg im Breisgau, 79104, Germany; 5Institute for Adaptive and Neural
Computation, School of Informatics, University of Edinburgh, EH8 9AB, UK
Correspondence: Philipp Weidel ‑ p.weidel@fz‑juelich.de
BMC Neuroscience 2016, 17(Suppl 1):P93
Perceptual decision-making is an intricate process implicating the
coordinated activity of multiple brain areas [1, 2]. Recent experimental studies demonstrate the existence of a complex interplay between
decision-related neural events and transient working memory processes [1], implemented by distributed circuits where specific subpopulations appear to be differentially involved in the evidence
accumulation process and subsequent behavioral outcomes [2]. This
results in observable divergences in choice-specific neuronal dynamics, unfolding as reproducible trajectories throughout the network’s
state-space [1] and hinting at the dissipative nature of the underlying dynamical system, which executes cognitively relevant processing
through transient trajectories.
Despite this evidence, the majority of modeling studies addressing
reward-modulated decision-making tend to simplify the formalization of environmental representations in the cortex as stable, attractor
states corresponding to discrete environmental states [3]. Even models
involving transient-based computations often simplify sensory stimuli
to a discrete set of inputs transduced as stochastic point processes [4].
These simplifications potentially draw an incomplete picture of neural
dynamics and therefore provide limited insights into the true nature of
computation in neural circuits.
References
1. Harvey CD, Coen P, Tank DW. Choice-specific sequences in parietal cortex
during a virtual-navigation decision task. Nature. 2012;484(7392):62–8.
2. Shadlen MN, Kiani R. Decision making as a window on cognition. Neuron.
2013;80(3):791–806.
3. Jitsev J, Morrison A, Tittgemeyer M Learning from positive and negative
rewards in a spiking neural network model of basal ganglia. In: The 2012
international joint conference on neural networks (IJCNN). IEEE; 2012.
4. Duarte R, Morrison A. Dynamic stability of sequential stimulus representations in adapting neuronal networks. Front Comput Neurosci.
2014;8(124).
5. Weidel P, Duarte R, Djurfeldt M, Morrison A. ROS-MUSIC toolchain (in
preparation).
6. Eliasmith C, Anderson CH. Neural engineering: computation, representation, and dynamics in neurobiological systems. MIT Press; 2004.
P94
Modulation of tuning induced by abrupt reduction of SST cell
activity
Jung H. Lee1, Ramakrishnan Iyer1, Stefan Mihalas1
1
Allen Institute for Brain Science, Seattle, WA 98109, USA
Correspondence: Jung H. Lee ‑ jungl@alleninstitute.org
BMC Neuroscience 2016, 17(Suppl 1):P94
Inhibitory interneurons have been considered pivotal in orchestrating
pyramidal neurons. Indeed, the optogenetic perturbation of inhibitory
cell types confirmed its validity. Recent studies [1, 2] have found that
the optogenetic stimulation of somatostatin positive (SST) interneurons, one of the three major inhibitory types, sharpens the tuning of
visual neurons, but its effect was conspicuous only when the optogenetic activation of SST cells was turned off abruptly. Specifically, with
4-s presentation of visual stimuli, the 1-s activation of SST cells resulted
in a sharper tuning, whereas 4-s activation did not induce significant
sharpening [2], which leads to a question: “Why does the length of
optogenetic stimulation render such a striking difference?” Lee et al.
suggested that the 1-s activation sharpens the tuning curve due to
the rebound activity of PV cells, and El-Boustani et al. suggested the
reduction of co-activation between PV and Pyr cell activity; see Ref. [2]
for the details.
In our study, we investigate the potential mechanisms underlying the
disparate effects between short and long activations of SST cells by
using the firing rate equations that expresses the interactions among
Pyr, SST and PV cells conveyed via cell-type specific connections
reported by Pfeffer et al. [3]. Our model consists of five populations:
BMC Neurosci 2016, 17(Suppl 1):54
Page 61 of 112
two pyramidal populations (Pyr1, 2), two PV cell populations (PV1, 2)
and SST cell population. We assume that Pyr1 and Pyr2 in close proximity respond to preferred and non-preferred stimuli, respectively.
The two pyramidal populations excite the shared SST cell population
which sends inhibition back to them. Since SST cells are known to be
connected to distant presynaptic pyramidal cells via long-horizontal
connections [4], the two SST populations in close proximity would
receive (almost) identical inputs, making the two SST populations
redundant. PV1 and PV2, which receive identical external background
inputs, interact with Pyr1 and Pyr2, respectively. Pyr1 and Pyr2 are not
directly connected, but they can indirectly interact with each other
through SST cell population.
Our model replicates the paradoxical finding that not 4-s activation of SST cells but 1-s activation leads to the sharper responses of
V1 neurons. In our model, PV cells provide synchronized inhibition to
pyramidal cells despite their distinctive receptive fields when SST cell
activation is abruptly turned off. If SST cells are stimulated during the
entire period of simulations (4 s), the induced synchronous inhibition from PV cells to pyramidal cells is not strong enough to induce
sharper responses. We also found that this synchronous inhibition can
be induced by the activation of VIP cells, raising the possibility that VIP
cells regulate V1 neural responses with the proposed mechanism.
the mouse running, as the enhanced VIP cell activity reduces surround
suppression.
Here we use a computational model of V1 consisting of multiple
cortical columns to address if VIP cell activation can enhance perception on moving objects. Our computational model is based on
an earlier multiple column model [6], and we refined it by incorporating VIP, SST and PV cells into the superficial layers of the
model. To build this refined model, we used two strategies. First,
we inferred the time course of synaptic events and the number of
connections from both experimental data [7] and parameters from
the earlier model. Second, we identified the minimal set of intercolumnar connections necessary for reproducing lateral interactions
observed in visual cortex [8]. To examine whether the enhanced VIP
cells could enhance responses to moving objects, we simulated a
single moving object by stimulating the columns sequentially in the
model. Our simulation results support our hypothesis: the column
responses during sequential stimulation increase as we increase
inputs to VIP cells.
Acknowledgements: We wish to thank the Allen Institute founders, Paul G. Allen and Jody Allen, for their vision, encouragement and
support.
References
1. Rudy B, Fishell G, Lee S, Hjerling-Leffler J. Three groups of interneurons
account for nearly 100 % of neocortical GABAergic neurons. Dev Neurobiol. 2011;71:45–61.
2. Jiang X, Shen S, Cadwelll C, Berence P, Sinz F, Ecker AS, et al. Principles of
connectivity among morphologically defined cell types in adult neocortex. Science. 2015;350:62–4.
3. Kepecs A, Fishell G. Interneuron cell types are fit to function. Nature.
2014;505:318–26.
4. Zhang S, Xu M, Kamigaki T, Hoang Do JP, Chang W-C, Jenvay S, et al.
Long-range and local circuits for top-down modulation of visual cortex
processing. Science. 2014;345:660–5.
5. Fu Y, Tucciarone JM, Espinosa JS, Sheng N, Darcy DP, Nicoll RA, et al. A cortical circuit for gain control by behavioral state. Cell. 2014;156:1139–52.
6. Wagatsuma N, Potjans TC, Diesmann M, Sakai K, Fukai T. Spatial and
feature-based attention in a layered cortical microcircuit model. PLoS
One. 2013;8:e80788.
7. Pfeffer CK, Xue M, He M, Huang ZJ, Scanziani M. Inhibition of inhibition
in visual cortex: the logic of connections between molecularly distinct
interneurons. Nat Neurosci. 2013;16:1068–76.
8. Xu S, Jiang W, Poo M-M, Dan Y. Activity recall in a visual cortical ensemble.
Nat Neurosci. 2012, 15:449–455.
References
1. Wilson NR, Runyan CA, Wang FL, Sur M. Division and subtraction by distinct cortical inhibitory networks in vivo. Nature. 2012;488(7411):343–8.
doi:10.1038/nature11347.
2. Lee S-H, Kwan AC, Dan Y. Interneuron subtypes and orientation tuning.
Nature. 2014;508(7494):E1–2. doi:10.1038/nature13128.
3. Pfeffer CK, Xue M, He M, Huang ZJ, Scanziani M. Inhibition of inhibition
in visual cortex: the logic of connections between molecularly distinct
interneurons. Nat Neurosci. 2013;16(8):1068–76. doi:10.1038/nn.3446.
4. Adesnik H, Bruns W, Taniguchi H, Huang ZJ, Scanziani M. A neural circuit
for spatial summation in visual cortex. Nature. 2012;490(7419):226–31.
doi:10.1038/nature11526.
P95
The functional role of VIP cell activation during locomotion
Jung H. Lee1, Ramakrishnan Iyer1, Christof Koch1, Stefan Mihalas1
1
Allen Institute for Brain Science, Seattle, WA 98109, USA
Correspondence: Jung H. Lee ‑ jungl@alleninstitute.org
BMC Neuroscience 2016, 17(Suppl 1):P95
Vasoactive intestinal polypeptide positive (VIP) inhibitory interneurons are commonly found in the superficial layers of cortices [1]. They
are distinct from other major cortical inhibitory cell types in terms of
connectivity and cellular mechanisms and exclusively inhibit somatostatin (SST) cells in visual cortex. This property is consistent with the
interneuron-selective interneuron group, which has been proposed
recently [2]. Indeed, bitufted cells in this group express VIP. Moreover,
VIP cells have nicotinic receptors rarely found in SST and parvalbumin
positive (PV) cells [3]. These recent studies lead to the hypothesis that
VIP cells play unique functions in cortical areas, which can be supported with evidence. The optogenetic activation of cingulate cortex
of mouse elicited a strong response in VIP cells of V1, suggesting the
central roles of VIP cells in top-down gain control [4]. Also, VIP cells are
nonspecifically depolarized when a mouse runs [5].
VIP cells disinhibit pyramidal cells by suppressing SST cell activity. That
is, when VIP cells are activated, pyramidal cell activity increases due
to reduction of inhibition from SST cells, which accounts for the gain
modulation. However, the advantage of VIP cell activation induced by
locomotion is not clear. We hypothesized that VIP cell activation leads
to better perception of moving objects since all visual objects would
appear to be in motion when a mouse runs. The strong surround suppression could prevent visual neurons from responding to those effective movements. In this sense, VIP cell activation may be beneficial to
Acknowledgements: We wish to thank the Allen Institute founders, Paul G. Allen and Jody Allen, for their vision, encouragement and
support.
P96
Stochastic inference with spiking neural networks
Mihai A. Petrovici1,†, Luziwei Leng1,†, Oliver Breitwieser1,†, David Stöckel1,†,
Ilja Bytschok1, Roman Martel1, Johannes Bill1, Johannes Schemmel1,
Karlheinz Meier1
Kirchhoff‑Institute for Physics, University of Heidelberg, Germany
Correspondence: Mihai A. Petrovici ‑ mpedro@kip.uni‑heidelberg.de †
Authors with equal contributions
BMC Neuroscience 2016, 17(Suppl 1):P96
Brains are adept at creating an impressively accurate internal model
of their surrounding based on incomplete and noisy sensory data.
Understanding this inferential prowess is not only interesting for neuroscience, but may also inspire computational architectures and algorithms for solving hard inference problems. Here, we give an overview
of our work on probabilistic inference with brain-inspired spiking networks, their advantages compared to classical neural networks and
their implementation in neuromorphic hardware.
In the neural sampling framework, we interpret spiking activity as sampling from distributions over binary random variables. By exploiting
the dynamics of spiking neurons with conductance-based synapses,
we have shown that their activation function can become symmetric in
the high-conductance state, which in turn enables Glauber-like dynamics in ensembles of noise-driven LIF networks [1, 2]. This allows the
BMC Neurosci 2016, 17(Suppl 1):54
straightforward construction of LIF networks that sample from previously
defined probability distributions.
When the parameters of the distribution are not well-defined, they need
to be learned from data. Due to their analogy to classical neural networks
such as Boltzmann machines, LIF networks are amenable to the same
learning algorithms and can be shown to match the performance of their
equally-sized abstract counterparts when trained on classical machinelearning datasets such as MNIST. However, spiking neural networks
endowed with short-term plasticity can travel more efficiently through
their associated state space, allowing them to simultaneously become
good generative and discriminative models of learned data, which is
notoriously difficult with conventional techniques such as Gibbs sampling. This finding points towards a distinct advantage of spike-based
computation and communication, which is relevant in any scenario
where spiking neural networks need to be able to escape local attractors.
This computational advantage of spiking sampling networks can be
further bolstered by emulation on an accelerated neuromorphic substrate. The core idea behind these devices is the direct emulation of
biological neuronal dynamics in VLSI circuits. Such hardware can far
surpass simulators running on conventional computing architectures
both in terms of speed and power consumption, but with the caveat of
having limited parameter precision, as well as other sources of disruptive noise [3, 4]. With some additional modifications, we have shown
how LIF networks can become robust to certain types of parameter
noise—both during training and during operation– thereby making
them amenable to a neuromorphic implementation with an acceleration factor of 104 compared to biological real-time.
An even more compelling argument for neuromorphic spike-based
inference can be made when considering that learning (in particular,
the simulation of synaptic plasticity) is by far the most time-consuming factor in simulations. In an effort to make expectation–maximization learning compatible with existing neuromorphic devices, we have
developed a network model that can use double-exponential STDP
with 4–6 bit weight resolution for learning and spike-based homeostasis for stabilization and robustness.
Page 62 of 112
(GCL). Co-stimulation of target retinal ganglion cells and overlying
axons results in irregular visual percepts, which can significantly limit
perceptual efficacy [1, 2]. This research explores how the characteristic
distributions of fiber orientation in different retinal layers result in differences between the activation of the axon initial segment and axons
of passage. Specifically, axons of passage of retinal ganglion cells are
characterized by a narrow distribution of fiber orientations, dominated
by the direction of passage towards the optic disk. In contrast, proximal axons in the GCL tend to have a wider distribution of orientations.
A model of extracellular stimulation that captures the effects of neurite orientation has been developed using a modified version of the
standard volume conductor model, known as the cellular composite
model [3], embedded in a four layer model of the retina. The cellular
composite model is used in this analysis as it addresses a number of
limitations of conventional volume conductor models and more accurately captures the spatiotemporal properties of neural tissue.
By generalizing the model to allow for analysis of fibers with arbitrary
orientations, simulations have been conducted to investigate the
interaction of neural tissue orientation, electrode placement, and stimulation pulse duration and amplitude.
Through an exhaustive parameter search, a set of stimulation pulse
durations, amplitudes and electrode positions are proposed to achieve
selective activation of axon initial segments. Using appropriate multiple electrode configurations and higher frequency stimulation, preferential activation of the axon initial segment is shown to be possible for
a range of realistic electrode-retina separation distances (Fig. 58).
These results establish a quantitative relationship between the timecourse of stimulation and physical properties of the tissue, such as
fiber orientation.
References
1. Fried SI, Lasker ACW, Desai NJ, Eddington DK. Axonal sodium-channel
bands shape the response to electric stimulation in retinal ganglion cells.
J Neurophysiol. 2009;101(4):1972–87.
This research was supported by EU grants #269921 (BrainScaleS),
#237955 (FACETS-ITN), #604102 (Human Brain Project) and the Manfred Stärk Foundation.
References
1. Petrovici MA, Bill J, Bytschok I, Schemmel J, Meier K. Stochastic inference
with deterministic spiking neurons. arXiv preprint arXiv:1311.3211, 2013.
2. Petrovici MA, Bytschok I, Bill J, Schemmel J, Meier K. The high-conductance state enables neural sampling in networks of LIF neurons. BMC
Neurosci. 20150;16(Suppl. 1):O2.
3. Petrovici MA, Vogginger B, Müller P, Breitwieser O, Lundqvist M, Muller L,
Ehrlich M, Destexhe A, Lansner A, Schüffny R, et al. Characterization and
compensation of network-level anomalies in mixed-signal neuromorphic
modeling platforms. PloS One. 2014;9(10):e108590.
4. Pfeil T, Grübl A, Jeltsch S, Müller E, Müller P, Petrovici MA, Schmuker M,
Brüderle D, Schemmel J, Meier K. Six networks on a universal neuromorphic computing substrate. Front Neurosci. 2013;7:11. ISSN 1662-453X.
doi:10.3389/fnins.2013.00011.
P97
Modeling orientation‑selective electrical stimulation with retinal
prostheses
Timothy B. Esler1, Anthony N. Burkitt1, David B. Grayden1, Robert R. Kerr2,
Bahman Tahayori3, Hamish Meffin4
1
NeuroEngineering Laboratory, Electrical & Electronic Engineering, The
University of Melbourne, Parkville VIC 3010, Australia; 2IBM Research,
Melbourne, Australia; 3Monash Institute of Medical Engineering, Monash
University, Melbourne, Australia; 4National Vision Research Institute,
Melbourne, Australia
Correspondence: Timothy B. Esler ‑ tesler@student.unimelb.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P97
One present challenge in electrical stimulation for epiretinal prostheses is how to avoid stimulating axons of passage in the nerve fiber
layer (NFL) that originate from distant regions of the ganglion cell layer
Fig. 58 A Simulation geometry showing the four modeled layers:
insulator (glass), vitreous, NFL, and GCL. Distance from membrane
threshold in mV for B parallel axons in a plane in the NFL and C
perpendicular axon initial segments in a plane in the GCL, when
stimulated with a 300 µs biphasic pulse with electrode-retina separation of 400 µm. Dotted contour marks the threshold level
BMC Neurosci 2016, 17(Suppl 1):54
2.
3.
Page 63 of 112
Rattay F, Resatz S. Effective electrode configuration for selective stimulation with inner eye prostheses. IEEE Trans Biomed Eng.
2004;51(9):1659–64.
Meffin H, Tahayori B, Sergeev EN, Mareels IMY, Grayden DB, Burkitt AN.
Modelling extracellular electrical stimulation: III. Derivation and interpretation of neural tissue equations. J Neural Eng. 2014;11(6):065004.
P98
Ion channel noise can explain firing correlation in auditory nerves
Bahar Moezzi1, Nicolangelo Iannella1,2, Mark D. McDonnell1
1
Computational and Theoretical Neuroscience Laboratory, School
of Information Technology and Mathematical Sciences, University
of South Australia, Australia; 2School of Mathematical Sciences, University
of Nottingham, Nottingham, UK
Correspondence: Bahar Moezzi ‑ bahar.moezzi@unisa.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P98
Neural spike trains are commonly characterized as a Poisson point
process. However, the Poisson assumption is a poor model for spiking
in auditory nerve fibers because it is known that interspike-intervals
display positive correlation over long time scales and negative correlation over shorter time scales. It has been suggested that ion channel opening and closing might not be well described by Markov
models. Instead, fractal ion channel gating could be used to take into
account the involvement of proteins in the conformational changes of
sub-states in the channel gating kinetics. Using a detailed biophysical model, we tested the hypothesis that fractal ion channel gating is
responsible for short and long term correlations in the auditory nerve
spike trains.
We developed a biophysical model based on the well-known Meddis model of the peripheral auditory system [1]. We introduced biophysically realistic ion channel noise to an inner hair cell membrane
potential model that includes (i) fractal fast potassium channels, (ii)
deterministic slow potassium channels, and (iii) a stochastic Markov
model for noisy calcium channels. We used Fano factor as a measure
of firing correlation.
We showed that the resulting simulated Fano factor time curves
have all the common attributes of the Fano factor of experimentally
recorded spike trains in the auditory nerve fibers, except the time scale
of corelation. Our model thus replicates macro-scale stochastic spiking
statistics in the auditory nerve fibers due to modeling stochasticity at
the micro-scale of potassium and calcium ion channels.
Reference
1. Meddis R. Auditory-nerve first-spike latency and auditory absolute
threshold: a computer model. J Acoust Soc Am. 2006;119(1):406–17.
P99
Limits of temporal encoding of thalamocortical inputs in a
neocortical microcircuit
Max Nolte1, Michael W. Reimann1, Eilif Muller1, Henry Markram1
1
Blue Brain Project, École Polytechnique fédérale de Lausanne (EPFL),
Geneva, Switzerland
Correspondence: Max Nolte ‑ max.nolte@epfl.ch
BMC Neuroscience 2016, 17(Suppl 1):P99
During naturalistic whisker motion, subsets of neurons in the same
barreloid of the rat ventroposterior medial thalamus (VPM) respond
synchronously with temporal precision to different kinetic features of
whisker movement (spike-time coding) [1]. Multiple synchronously
firing VPM cells can trigger temporally precise responses in the somatosensory cortex, such as those observed during full whisker deflection or active touch, but the minimum number of synchronously firing
VPM cells needed to reliably drive the spiking of cortical cells is not
known.
In this study, we use the Blue Brain Project’s digital reconstruction of
a somatosensory microcircuit of a juvenile rat [2] to characterize how
many synchronously firing VPM cells are needed to reliably drive individual cells of different morphological types in the rat somatosensory
Fig. 59 A Mean spike-timing reliability (similar correlation-based
measure as in [3], but with firing rate adaption). The reliability of the
VPM input is 0.55. B Mean probability of firing within 2–12 ms after
the initial input VPM spike in each trial. C Mean ratio of spikes occurring within 2–12 ms after a VPM spike, out of all spikes. Mean of 30
(L3/4 excitatory), 50 (L5/6 exc.), 40 (L3/4 inhibitory) and 30 cells (L5/6
inh.) respectively
cortex. We activate an increasing number of synchronously firing VPM
fibers (with in vivo VPM spike trains from experiments published in
[1]) in both simulations of single cells, and simulations of the whole
reconstructed microcircuit with only a small number of active VPM fibers. We find that inhibitory neurons in layers 3 and 4 quickly approach
maximum spike-timing reliability when receiving input from 10 to 15
synchronously firing VPM neurons. Excitatory neurons in layers 3 and 4
require substantially more synchronous VPM fibers, but less than excitatory neurons in layers 5 and 6 (see Fig. 59). With an average of eight
synapses per connection, these numbers are significantly higher than
what has been observed in a previous in silico study in the cat visual
cortex [3]. In addition to the difference in animal and sensory system,
we show that this decrease of reliability can be partly explained by
a lower synaptic release probability in vivo than in vitro caused by a
lower extracellular calcium concentration in vivo [4], which is taken
into account in our simulations [2].
Finally, we describe how the requirement for synchronous, redundant
VPM inputs limits the maximum amount of asynchronous, temporally
precise VPM activity (in subsets of synchronous VPM neurons) that can
be reliably encoded in a neocortical microcircuit.
Acknowledgements: This work was supported by funding from the
ETH Domain for the Blue Brain Project (BBP). The BlueBrain IV IBM
BlueGene/Q system is financed by ETH Board Funding to the Blue
Brain Project and hosted at the Swiss National Supercomputing Center
(CSCS). We thank M. Bale and R. Petersen for providing the VPM spike
trains.
References
1. Bale MR, Ince RAA, Santagata G, Petersen RS. Efficient population coding
of naturalistic whisker motion in the ventro-posterior medial thalamus
based on precise spike timing. Front Neural Circuits. 2015;50.
2. Markram H, et al. Reconstruction and simulation of neocortical microcircuitry. Cell. 2015;163(2):456–92.
3. Wang H, Spencer D, Fellous J, Sejnowski T. Synchrony of thalamocortical
inputs maximizes cortical reliability. Science. 2010;328:106–9.
4. Borst JGG. The low synaptic release probability in vivo. Trends Neurosci.
2010;33(6):259–66.
P100
On the representation of arm reaching movements: a
computational model
Antonio Parziale1, Rosa Senatore, Angelo Marcelli1
Department of Information and Electrical Engineering, University
of Salerno, 84084, Fisciano (SA), Italy
Correspondence: Antonio Parziale ‑ anparziale@unisa.it
BMC Neuroscience 2016, 17(Suppl 1):P100
Experimental studies on the spinal cord (SC) have shown that SC is
not a simple relay station for transmitting information to and from
supraspinal centers but “it is a highly evolved and complex part
BMC Neurosci 2016, 17(Suppl 1):54
of the CNS that has considerable computational ability” [1]. Limb
movements are planned and initiated by the brain but they cannot
be performed without a spinal cord and the intricate feedback systems that reside within it [2]. In the last years, computational models
have been devised in order to explain the role of the spinal cord in
the translation from motor intention to motor execution [3], in sensorimotor control and learning of movements [4], in investigating
how the supraspinal centers can control the cord [5], for providing
evidence that CNS can plan and control movements without a representation of complex bodily dynamics because the creation and
coordination of dynamic muscle forces is entrusted to the spinal
feedback mechanisms [6], for investigating how the central nervous
system coordinates the activation of both α and γ motoneurons during movement and posture [7].
Here we propose a computational model of the local interneuron
networks within SC to evaluate how spinal and supraspinal centers
can interact for performing a movement. We model a one-degree of
freedom system representing an arm learning and executing reaching movements. The model incorporates the key anatomical and
physiological features of the neurons in SC, namely interneurons Ia, Ib
and PN and Renshaw cells, and their interconnections [2]. The model
envisages descending inputs coming from both rostral and caudal M1
motor cortex and cerebellum (through the rubro- and reticulo-spinal
tracts), local inputs from both Golgi tendon organs and spindles, and
its output is directed towards α motoneurons, which also receive
descending inputs from the cortex and local inputs from spindles.
The model envisages virtual muscle [8] for modeling musculoskeletal
mechanics and proprioceptors.
Our simulations show that the CNS may produce elbow flexion movements with different properties by adopting different strategies for the
recruitment and the modulation of interneurons and motoneurons.
One interesting results is that the speed-accuracy tradeoff predicted
by the Fitts’ law [9] does not follow from the structure of the system,
that is capable of performing fast and precise movements, but arises
from the strategy adopted to produce faster movements, by starting
from a pre-learned set of motor commands useful to reach the target position and by modifying only the activations of the PN and α
neurons.
Other simulations show that when a suddenly variation of the target
position happens after the onset of a learned movement, the descending inputs from the cerebellum can be exploited for the online correction of the movement trajectory by regulating the activity of PN cells.
This result agrees with the experimental studies suggesting that the
CNS modulates interneurons networks to execute a visually guided
online correction.
References
1. Burke RE. Spinal cord. Scholarpedia. 2008;3(4):1925.
2. Pierrot-Deseilligny E, Burke DJ. The circuitry of the human spinal cord:
neuroplasticity and corticospinal mechanisms. Cambridge: Cambridge
University Press; 2012.
3. Bullock D, Grossberg S. VITE and FLETE: Neural modules for trajectory
formation and tension control. Volitional Action. 1989;253–97.
4. Tsianos GA, Goodnes J, Loeb GE. Useful properties of spinal circuits for
learning and performing planar reaches. J Neural Eng. 2014;11:1–21.
5. Raphael G, Tsianos GA, Loeb GE. Spinal-like regulator facilitates control of
a two-degree-of-freedom wrist. J Neurosci. 2010;30:9431–44.
6. Buhrmann T, Di Paolo EA. Spinal circuits can accommodate interaction
torques during multijoint limb movements. Front Comput Neurosci.
2014;8:1–18.
7. Li S, Hao M, He X, Marquez JC, Niu CM, Lan N. Coordinated alpha and
gamma control of muscles and spindles in movement and posture. Front
Comput Neurosci. 2015;9:1–15.
8. Cheng EJ, Brown IE, Loeb GE: Virtual Muscle: a computational approach
to understanding the effects of muscles properties on motor control. J
Neurosci Methods. 2000;101:117–30.
9. Fitts PM. The information capacity of the human motor system in controlling the amplitude of movement. J Exp Psychol. 1954;47(6):381–91.
Page 64 of 112
P101
A computational model for investigating the role of cerebellum
in acquisition and retention of motor behavior
Rosa Senatore1,2, Antonio Parziale1, Angelo Marcelli1
1
Department of Information and Electrical Engineering and Applied
Mathematics, University of Salerno, Fisciano (SA), 81100, Italy; 2Laboratory
of Neural Computation, Istituto Italiano di Tecnologia, Rovereto (TN),
38068, Italy
Correspondence: Rosa Senatore ‑ rsenatore@unisa.it
BMC Neuroscience 2016, 17(Suppl 1):P101
Experimental studies on the cerebellum (CB) have provided a large
body of knowledge about its anatomical and physiological features,
the neural processes and the phenomena of synaptic plasticity occurring within both the cerebellar cortex and nuclei [1]. The emerging
picture is that the CB plays a crucial role in the acquisition and/or
retention of motor behaviors and is involved in several cognitive functions [2, 3], therefore several CB models and simulations of its neural
processes have been proposed [3–5]. Here we investigated, through
a modeling approach, the role of the CB in three different behaviors:
vestibulo-ocular reflex (VOR) adaptation, motor learning, and eyeblink
conditioning. Different cerebellar areas are involved in these functions:
the control of the amplitude and timing of the VOR involves the vestibulocerebellum, learning novel limb movements involves the lateral cerebellar cortex and its connections to the dentate nucleus and
acquisition of the eyeblink conditioned responses involves cerebellar
cortex areas (lobule HVI) connected to the interposed nucleus[1, 3].
It is noteworthy that the CB is characterized by the remarkable regularity and geometrical structure of its circuits: cerebellar neurons are
arranged in a highly regular manner as repeating units, the cerebellar microcomplexes [1]. Therefore the uniform structure of the CB and
the contribution of different cerebellar areas to specific behaviors raise
the possibility that different behaviors are based on a common ‘neural
computation within the cerebellum’.
We developed a model (using the Leabra framework in emergent neural simulation software [6]) that incorporates the key anatomical and
physiological features of the cerebellar microcomplex, whose behavior was analyzed for investigating the neural processes occurring during the acquisition of novel motor behaviors, classically conditioned
responses and VOR adaptation. Since the neural circuits involved in
these behaviors present some differences, in terms of the input/output areas sending signals to or receiving signals from the CB, we developed three models, which share the same core network, made up of
a set of cerebellar microcomplexes (comprising cerebellar cortex neurons and their connections to nuclear and olivary neurons), but which
include different anatomical connections from/to different extra cerebellar regions: (a) “VOR model”, comprising the vestibulocerebellum
(flocculus and vestibular nucleus) and its connections with the dorsal cap region of the inferior olive and oculomotor nuclei; (b) “Motor
model”, comprising the lateral cerebellum (lateral cerebellar cortex
and dentate nucleus) and its anatomical connections with the inferior
olive, thalamus and motor cortex; (c) “Conditioning model”, comprising the lobule HVI of the cerebellar cortex and its connections to the
interposed nucleus, and their external connections with the dorsal
accessory olive, red nucleus and oculomotor nuclei.
Our simulations suggest that the CB performs the same computational operation on whichever afferent information it receives, that
the appearance of the ‘teaching’ signal conveyed by the climbing fibers could be the explanation for functional differentiation and that
different types and sites of synaptic plasticity are involved in different
behaviors.
References
1. Ghez C, Thach WT. The cerebellum. In: Kandel ER, Schwartz JH, Jessell TM,
editors. Principles of neural science. McGraw-Hill; 2000. p. 832–52.
2. Koziol LF, Budding D, Andreasen N, D’Arrigo S, Bulgheroni S, Imamizu H,
Ito M, Manto M, Marvel C, Parker K, et al. Consensus paper:the cerebellum’s role in movement and cognition. Cerebellum. 2014;13(1):151–77.
BMC Neurosci 2016, 17(Suppl 1):54
3.
4.
5.
6.
Manto M, Bower JM, Conforto AB, Delgado-Garcia JM, da Guarda SN, Gerwig M, Habas C, Hagura N, Ivry RB, Marien P, et al. Consensus paper: roles
of the cerebellum in motor control–the diversity of ideas on cerebellar
involvement in movement. Cerebellum. 2012;11(2):457–87.
Houk JC, Buckingham JT, Barto AG. Models of the cerebellum and motor
learning. Behav Brain Sci. 1996;19(3):368–83.
Medina JF, Mauk MD. Computer simulation of cerebellar information
processing. Nat Neurosci. 2000;3(Suppl.):1205–11.
Aisa B, Mingus B, O’Reilly R. The emergent neural modeling system.
Neural Netw. 2008;21(8):1146–52.
P102
The emergence of semantic categories from a large‑scale brain
network of semantic knowledge
K. Skiker1, M. Maouene2
1
LIST Laboratory, FST, Abdelmalek Essaadi’s University, Tangier, Morocco;
2
Department of computer science, ENSAT, Abdelmalek Essaadi’s
University, Tangier, Morocco
Correspondence: K. Skiker ‑ skiker.kaoutar85@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P102
In cognitive neuroscience, the issue of how semantic categories (e.g.
animals, tools, fruits/vegetables) are organized in the brain is still
debated (Caramazza and Mahon 2006). Some authors postulate that
semantic categories are explicitly represented in specific brain areas
developed through evolutionary pressure for rapid classification and
categorization of animals, tools and foods (Caramazza and Shelton
1998). Other researches argue that semantic categories are not explicitly represented, instead emerge from distributed semantic knowledge
(Martin 2007; Tyler and Moss 2001). However, a little is known about
how semantic knowledge is structured within the brain for fast and
efficient emergence of semantic categories. In this paper, we hypothesize that semantic knowledge is supported by a large-scale brain
network that shows the properties of segregation and integration. To
test this hypothesis, we first examine where semantic knowledge is
nested in the brain; we present functional neuroimaging studies suggesting that semantic knowledge (e.g. visual, auditory, tactile; action,
olfactory/gustatory) is grounded in modality specific association brain
areas (e.g. visual association areas, auditory association areas, somatosensory association areas) (Barsalou 2008; Goldberg et al. 2006).
Then, we derive the connectivity between brain areas where semantic
knowledge is nested from Hagmann’s connectivity matrix (Hagmann
et al. 2008) freely available. Finally, we examine the properties of the
connectivity matrix using graph measures including clustering coefficient and characteristic path length. Our findings show that a largescale brain network of features exhibit the small world property with
high clustering coefficient (C = 0.48) and low path length (L = 2.49).
These properties indicate a balance between segregation (high clustering) and integration (low path length) that are essential for the
fast and efficient emergence of semantic categories from distributed
semantic knowledge.
References
1. Barsalou, Lawrence W. Grounded cognition. Annu Rev Psychol. 2008;59:
617–45.
2. Caramazza A, Mahon BZ. The organisation of conceptual knowledge in
the brain: the future’s past and some future directions. Cogn Neuropsychol. 2006;23(1):13–38.
3. Caramazza A, Shelton JR. Domain-specific knowledge systems
in the brain the animate-inanimate distinction. J Cogn Neurosci.
1998;10(1):1–34.
4. Goldberg RF, Perfetti CA, Schneider W. Perceptual knowledge retrieval
activates sensory brain regions. J Neurosci. 2006;26(18):4917–21.
5. Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey CJ, Wedeen VJ,
Sporns O. Mapping the structural core of human cerebral cortex. PLoS
Biol. 2008;6(7): e159.
6. Martin A. The representation of object concepts in the brain. Annu Rev
Psychol. 2007;58:25–45.
7. Tyler LK, Moss HE. Towards a distributed account of conceptual knowledge. Trends Cogn Sci. 2001;5(6):244–52.
Page 65 of 112
P103
Multiscale modeling of M1
multitarget pharmacotherapy for dystonia
Samuel A. Neymotin1,2, Salvador Dura‑Bernal1, Alexandra Seidenstein1,3,
Peter Lakatos4, Terence D Sanger5,6, and William W Lytton1,7
1
Department Physiology & Pharmacology, SUNY Downstate, Brooklyn,
NY 11203, USA; 2Department Neuroscience, Yale University School
of Medicine, New Haven, CT, USA; 3Departmentt of Chemical
and Biomedical Engineering, Tandon School of Engineering, NYU,
Brooklyn, NY, USA; 4Nathan Kline Institute for Psychiatric Research,
Orangeburg, NY, USA; 5Department Biomedical Engineering, University
of Southern California, Los Angeles, CA, USA; 6Div Neurology, Child
Neurology and Movement Disorders, Children’s Hospital Los Angeles,
LA, CA, USA; 7Department Neurology, Kings County Hospital Center,
Brooklyn, NY 11203, USA
Correspondence: Samuel A. Neymotin ‑ samn@neurosim.downstate.edu
BMC Neuroscience 2016, 17(Suppl 1):P103
Dystonia is a movement disorder that produces involuntary sustained
muscle contractions. Different types of dystonia likely involve primary
or induced pathologies across multiple brain areas including basal
ganglia, thalamus, cerebellum, and sensory and motor cortices. Due
to lack of therapeutic alternatives, much current treatment involves
paralyzing affected muscles directly with painful injections of botulinum toxin. Primary motor cortex (M1) represents a potential target for
therapy. M1 pathological dynamics in some forms of dystonia include
hyperexcitability and altered beta oscillations. In order to further
develop understanding of motor cortex involvement in this disease
and to look at potential drug cocktails (multitarget polypharmacy),
we developed a multiscale model of M1 across spatial scales, ranging from molecular interactions, up to cellular and network levels. The
model contains 1715 compartmental model neurons with multiple
ion channels and intracellular molecular dynamics [1, 2]. Wiring and
arrangements of cellular layers of the model was based on previously
recorded electrophysiological data obtained from mouse M1 circuit
mapping experiments. Simulations were run in the NEURON simulator and intracellular dynamics utilized the reaction–diffusion module
[3]. The chemophysiological component of the simulation focused on
calcium (Ca) handling, and Ca regulation of hyperpolarization-activated cyclic nucleotide-gated (HCN) channels. The Ca signaling was
modeled in conjunction with intracellular cytosolic and endoplasmic
reticulum (ER) volumes, inositol triphosphate (IP3) production via a
metabotropic glutamate receptor signaling cascade, and ER IP3 and
ryanodine receptors (RYR) which release ER Ca into the cytosol. The
model reproduced the pathological dynamics providing hyperexcitability and synchronous beta oscillations across cortical layers. We
applied independent random variations to multiple ion channel densities (multiple cell membrane channels: HCN, channels for Na, K, Ca;
RYR, IP3 channels in ER), to identify pathological and physiological
simulation sets. Experiments with these models demonstrated degeneracy, with multiple routes that produced the pathological syndrome.
In most cases, there was no single parameter alteration which would
induce the change from pathological to physiological dynamics. We
used support vector machines to assess the high dimensional parameter space to provide overall direction for passage from an overall
pathological to an overall physiological region of parameter space,
enabling prediction of multitarget drug cocktails that would be likely
to move the system from dystonic to physiological dynamics.
Acknowledgements: Research supported by NIH grant R01
MH086638, NIH grant U01 EB017695, NIH grant R01 NS064046, NIH
grant R01 DC012947.
References
1. Neymotin SA, McDougal RA, Bulanova AS, Zeki M, Lakatos P, Terman D,
Hines ML, Lytton WW. Calcium regulation of HCN channels supports
persistent activity in a multiscale model of neocortex. Neuroscience.
2016;316:344–66.
2. Neymotin SA, McDougal RA, Sherif MA, Fall CP, Hines ML, Lytton
WW. Neuronal calcium wave propagation varies with changes in
BMC Neurosci 2016, 17(Suppl 1):54
3.
endoplasmic reticulum parameters: a computer model. Neural Comput.
2015;27:898–924.
McDougal RA, Hines ML, Lytton WW. Reaction–diffusion in the NEURON
simulator. Front Neuroinform. 2013;7:28.
P104
Effect of network size on computational capacity
Salvador Dura‑Bernal1, Rosemary J. Menzies2, Campbell McLauchlan2,
Sacha J. van Albada3, David J. Kedziora2, Samuel Neymotin1, William W.
Lytton1, Cliff C. Kerr2
1
Department of Physiology & Pharmacology, SUNY Downstate Medical
Center, Brooklyn, NY 11023, USA; 2Complex Systems Group, School
of Physics, University of Sydney, Sydney, NSW 2006, Australia; 3Institute
of Neuroscience and Medicine (INM‑6), Jülich Research Centre and JARA,
Jülich, Germany
Correspondence: Cliff C. Kerr ‑ cliff@thekerrlab.com
BMC Neuroscience 2016, 17(Suppl 1):P104
There is exceptionally strong circumstantial evidence that organisms
with larger nervous systems are capable of performing more complex
computational tasks. Yet relatively few studies have investigated this
effect directly, instead typically treating network size as a fixed property of a simulation while exploring the effects of other parameters.
Recently, Diehl and Cook [1] found that network performance did
increase modestly with network size; however, larger networks also
required longer training times to achieve a given performance. In this
work, we directly addresses the relationship between network size and
computational capacity by using a biomimetic spiking network model
of motor cortex to direct a virtual arm towards a target via reinforcement learning [2]. The reaching task was performed by a two-joint
virtual arm controlled by four muscles (flexor and extensor muscles
for shoulder and elbow joints). These muscles were controlled by a
neural model that consisted of excitatory and inhibitory Izhikevich
neurons in three cortical populations: a proprioceptive population,
which received input from the current arm position; a motor population, which was used to drive the arm muscles; and a sensory population, which served as the link between the proprioceptive and motor
populations. The model was trained to reach the target using exploratory movements coupled with reinforcement learning and spike-timing dependent plasticity (STDP). The model was implemented using
NEURON.
A major challenge in scaling network size is that not all properties of
the network can be held constant. As shown by van Albada et al. [3],
while first-order properties (such as average firing rate) can be maintained, there are limitations in preserving second- and higher-order
statistical properties (such as noise correlations). Thus, we explored
multiple different ways of scaling the connectivity of the network,
including (a) preserving connection probability, scaling connection
weight to be inversely proportional to model size, and increasing
the variance of the external drive; and (b) reducing connection probability to preserve average node degree and leaving other parameters
unchanged. In addition, we explored scaling each of the neuronal
population groups versus only scaling the sensory (processing) population group. Large differences were observed in network dynamics
and statistics based on different scaling choices. However, the relationship between network size and task performance was significant
only for certain specific choices of model parameters. Overall, task
performance is highly sensitive to the network’s metaparameters, such
as STDP learning rates. We found that these must be optimized specifically for different network sizes; otherwise, differences in suitability of these parameters overwhelm the intrinsic advantages of larger
networks. In conclusion, while network size does affect computational
capacity, the relationship is strongly dependent on the manner in
which the scaling is implemented.
References
1. Diehl PU, Cook M. Unsupervised learning of digit recognition using spiketiming-dependent plasticity. Front Comp Neurosci. 2015;9:99.
2. Dura-Bernal S, Li K, Neymotin SA, Francis JT, Principe JC, Lytton WW.
Restoring behavior via inverse neurocontroller in a lesioned cortical spiking model driving a virtual arm. Front Neurosci. 2016;10:28.
Page 66 of 112
3.
Van Albada SJ, Helias M, Diesmann M. Scalability of asynchronous networks is limited by one-to-one mapping between effective connectivity
and correlations. PLoS Comput Biol. 2015;11:e1004490.
P105
NetPyNE: a Python package for NEURON to facilitate
development and parallel simulation of biological neuronal
networks
Salvador Dura‑Bernal1, Benjamin A. Suter2, Samuel A. Neymotin1, Cliff
C. Kerr3, Adrian Quintana4, Padraig Gleeson4, Gordon M. G. Shepherd2,
William W. Lytton1
1
Department Physiology & Pharmacology, SUNY Downstate, Brooklyn, NY
11203, USA; 2Department Physiology, Northwestern University, Chicago, IL
60611, USA; 3Complex Systems Group, School of Physics, University of Sydney,
Sydney, NSW 2006, Australia; 4Department of Neuroscience, Physiology &
Pharmacology, University College London, London WC1E6BT, UK
Correspondence: Salvador Dura‑Bernal ‑ salvadordura@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P105
NEURON is a widely used neuronal simulator, with over 1600 published models. It enables multiscale simulation ranging from the
molecular to the network level. However, learning to use NEURON,
especially running parallel simulations, requires much technical training. NetPyNE (Network development Python package for NEURON)
greatly facilitates the development and parallel simulation of biological neuronal networks in NEURON, potentially bringing its benefits to
a wider audience, including experimentalists. It is also intended for
experienced modelers, providing powerful features to incorporate
complex anatomical and physiological data into models.
NetPyNE seamlessly converts a set of high-level specifications into a
NEURON model. Specifications are provided in a simple, standardized,
declarative format, based solely on Python’s lists and dictionaries. The
user can define network populations and their properties, including
cell type, number or density. For each cell type, the user can define
morphology, biophysics and implementation, or choose to import
these from existing files (HOC templates or Python classes). Cell models for each population can be easily changed, and several models can
be combined to generate efficient hybrid networks, e.g. composed of
Hodgkin–Huxley multicompartment cells and Izhikevich point neurons. NetPyNE provides an extremely flexible format to specify connectivity, with rules based on pre- and post-synaptic cell properties, such
as cell type or location. Multiple connectivity functions are available,
including all-to-all, probabilistic, convergent or divergent. Additionally, connectivity parameters (e.g. weight, probability or delay) can
be specified as a function of pre/post-synaptic spatial properties. This
enables implementation of complex biological patterns, such as delays
or connection probabilities that depend on distance between cells, or
weights that depend on the post-synaptic neuron’s cortical depth. The
subcellular distribution of synapses along the dendrites can be specified, and is automatically adapted to the morphology of each model
neuron. Learning mechanisms, including spike-timing dependent plasticity and reinforcement learning, can be readily incorporated.
Using the high-level network specifications, NetPyNE instantiates the
full model (all cells and connections) as a hierarchical Python structure
including the NEURON objects necessary for simulation. Based on a set
of simulation options (e.g. duration, integration step), NetPyNE runs
the model in parallel using MPI, eliminating the burdensome task of
manually distributing the workload and gathering data across computing nodes. Optionally NetPyNE plots output data, such as spike
raster plots, LFP power spectra, connectivity matrix, or intrinsic timevarying variables (e.g. voltage) of any subset of cells. To facilitate data
sharing, the package saves and loads the high-level specifications,
instantiated network, and simulation results using common file formats (Pickle, Matlab, JSON or HDF5). NetPyNE can convert instantiated
networks to and from NeuroML, a standard data format for exchanging models in computational neuroscience.
NetPyNE has been used to develop a variety of multiscale models: primary motor cortex with cortical depth-dependent connectivity; the
claustrum; and sensorimotor cortex that learns to control a virtual arm.
The package is open source, easily installed, and includes comprehensive online documentation, a step-by-step tutorial and example
BMC Neurosci 2016, 17(Suppl 1):54
networks (www.neurosimlab.org/netpyne). We believe this tool will
strengthen the neuroscience community and encourage collaborations between experimentalists and modelers.
Acknowledgements: Research supported by NIH grant U01 EB017695
and DARPA grant N66001-10-C-2008.
P107
Inter‑areal and inter‑regional inhomogeneity in co‑axial
anisotropy of Cortical Point Spread in human visual areas
Juhyoung Ryu1, Sang‑Hun Lee1
1
Brain and Cognitive Science, Seoul National University, Seoul 151‑742,
Republic of Korea
Correspondence: Juhyoung Ryu ‑ jh67753737@snu.ac.kr, visionsl@snu.
ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P107
A focal visual stimulus can evoke widespread neural activation far
beyond the directly stimulated site, a phenomenon referred to as the
“cortical point spread” (CPS). The lateral connections among neurons
in the early visual cortex have been proposed as a likely anatomical conduit for the CPS, and recent functional studies on humans [1]
and non-human primates [2] demonstrated that the CPS is spatially
anisotropic, spread preferentially along with the axis of stimulus orientation, dubbed as ‘coaxial anisotropy.’ Although these two seminal
studies documented the coaxial anisotropy robustly in two different
species, there are several remaining questions to be further explored.
First, previous human psychophysical studies reported substantial
degrees of inhomogeneity in association field characteristics over the
visual space (e.g., crowding effects), which implies the presence of corresponding inhomogeneity of coaxial anisotropy. Second, the animal
study [2] examined the coaxial anisotropy only from V1 of two monkeys, and the human study [1] reported a substantial degree of individual differences and inter-areal differences. The current study is set
out to address these two aspects of coaxially anisotropic CPS.
We acquired time series of functional magnetic resonance imaging
(fMRI) measurements in V1 while human individuals viewed a ring or
wedge of Gabor patches that slowly drifted along the radial or tangential axis over a spatially extended (up to 8° in radius) region of
retinotopic space (Fig. 60A). The orthogonal combination of two different drifting direction and stimulus orientation generated two interesting viewing conditions: coaxial and orthoaxial conditions (boxed
and unboxed panels, respectively, in Fig. 60A). For individual gray
matter units (2 mm iso volume voxels) in the early visual cortex (V1,
V2, V3), we quantified the degree and sign of coaxial anisotropy by
comparing the width of fMRI response profiles between the coaxial
and orthoaxial conditions. In specific, we first estimated the width of
CPS at the half of its maximum response respectively for two viewing
conditions − coaxial condition (Wc) and orthoaxial condition (Wo),
Fig. 60 Stimuli and fMRI results. A The snapshot of traveling Gabors
are shown for the four different conditions. The black arrows represent
a moving direction of wedge or ring. B Significant (yellow, t test
p < 0.001) coaxial anisotropy in all subjects. C Coaxial anisotropy
across visual areas (V1, V2, V3)
Page 67 of 112
then computed coaxial anisotropy index by taking the singed contrast
between these two width estimates: CAI = (Wc − Wo)/(Wc + Wo).
Results The results replicated those in the previous study [1]: in all
of the subjects inspected, the width of CPS was significantly greater
along the coaxial axis than along the orthoaxial axis (Student’s t test,
p < 0.001), and the CAIs ranged from +0.05 to +0.15 (Fig. 60B). In addition, we found two interesting new findings: first, coaxial anisotropy
tended to decrease along the processing hierarchy (V1 > V2 > V3;
Fig. 60C); second, coaxial anisotropy tended to be more pronounced
along the cardinal axes (horizontal meridians in particular) in retinotopic space.
References
1. Park SH, Cha K, Lee SH. Coaxial anisotropy of cortical point spread in
human visual areas. J Neurosci. 2013; 33(3):1143–56a.
2. Michel MM, Chen Y, Geisler WS, Seidemann E. An illusion predicted by V1
population activity implicates cortical topography in shape perception.
Nat Neurosci. 2013; 16(10):1477–83.
P108
Two bayesian quanta of uncertainty explain the temporal
dynamics of cortical activity in the non‑sensory areas
during bistable perception
Joonwon Lee1, Sang‑Hun Lee1
1
Department of Brain and Cognitive Sciences, Seoul National University,
Seoul 151‑742, Korea
Correspondence: Joonwon Lee ‑ jwl89@snu.ac.kr; visionsl@snu.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P108
Bistable perception—involuntary fluctuation over time in perceptual
appearance despite unchanging physical stimulation on sensory organs—
has been a popular tool for exploring neural loci substantiating transitions in
perceptual awareness. To track observers’ ongoing percepts, experimenters
make the observers actively report binary states of their perception. Unfortunately, observers’ engagement in perceptual tracking is likely to invite the
brain activities that are not associated directly with perceptual transition
per se, but rather reflect the cognitive processes ensuing from temporal
sequences of perceptual transitions. We reasoned that, in a computational
perspective, the latter, post-transition component of bistable perception is
majorly driven by two separate quanta of uncertainty: unexpected (UU) and
expected uncertainty (EU), which have been proposed to relate to neuromodulatory systems of norepinephrine and acetylcholine, respectively [1].
We found the Dynamic Belief Model (DBM) [2] particularly relevant to the
concurrent measurement of those two quanta of uncertainty (i) because
it is capable of updating the moment-to-moment, expected probability
of binary events based on a recent history of those events (Fig. 61A) and
(ii) because this probability directly estimates EU whereas the disparity
between the expected probability and perceived outcome quantize UU.
Fig. 61 Bayesian estimation to predict BOLD dynamics around
switch. A Bayesian inference model of iteratively updated prior, input
likelihood, and combined posterior. B (Upper) Uncertainty-driven
BOLD estimated from Bayesian model locked to transition under
different duration conditions. Time-series is built purely from real
behavior history. (Lower) Average % BOLD signal of ACC region in 8
subjects (14 sessions)
BMC Neurosci 2016, 17(Suppl 1):54
With the DBM in hand, we predicted the time courses of UU and EU
(red curves in Fig. 61B), and explored the cortical loci substantiating
those two kinds of uncertainty by acquiring fMRI measurements while
human observers viewed a ‘structure-from-motion (SfM)’ display, in
which ambiguous 2D motion of coherently moving dots gives perceptual alternations in 3D motion perception between bistable states,
clockwise vs counterclockwise rotational motion. To compensate for
the temporal resolution of fMRI activity, we slowed down the dynamcis of bistable perception using the intermittent stimulation technique, which allowed us to identify gray-matter units (voxels) whose
variability in fMRI time course can be explained by UU or EU.
As expected from previous studies, cortical activity increased substantially during the transition periods in many distributed brain regions.
More importantly, the fMRI time series in these transition-locked
regions were explained by the weighted linear sum of the time series
of UU and EU quantity, some exhibiting greater weights for UU and
others greater weights for EU. In additions, the time series of pupil size
of the observers resembled the predicted time courses of UU, consistent with the previously reported tight linkage between UU and
the LC-NE system. We conclude that the cortical activities previously
claimed as being responsible for triggering perceptual transition are
likely to reflect two post-transition cognitive quanta of uncertainty.
References
1. Yu AJ, Dayan P. Uncertainty, neuromodulation, and attention. Neuron.
2005;46:681–92.
2. Yu AJ, Cohen JD. Sequential effects: superstition or rational behavior?
NIPS. 2009;21:1873–80.
P109
Optimal and suboptimal integration of sensory and value
information in perceptual decision making
Hyang Jung Lee1, Sang‑Hun Lee1
1
Department of Brain and Cognitive Neuroscience, Seoul National
University, Gwanak‑gu, South Korea
Correspondence: Hyang Jung Lee ‑ hyangjung.lee@snu.ac.kr, visionsl@
snu.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P109
Optimization of decision-making often requires the effective integration of sensory and value information, particularly when sensory
inputs are ambiguous and the criterion for the successful decision
changes stochastically over time (e.g., a hitter making ‘strike’ or ‘ball’
decisions by integrating visual information and an umpire’s calls).
We adapted this ‘sensory-value integration’ situation to a laboratory,
where 30 human subjects classified ring stimuli into ‘small’ or ‘large’
based on the perceived ring size and the trial-by-trial feedback (‘correct’ or ‘incorrect’) for judgment. The key manipulation was, unbeknownst to subjects, to induce a slight amount of bias, favoring either
‘small’ or ‘large’ choices, or staying ‘unbiased’. Inspired by previous animal studies (Corrado et al. 2005; Lau and Glimcher 2005; Busse et al.
2011), we developed a Linear-Nonlinear-Poisson model to describe
the dynamics of how humans adapt their moment-to-moment perceptual decision to subtle, yet volatile environmental feedbacks. The
integration process takes place at the Linear stage where a decision
variable is formed by combining the sensory information in the current trial and the reward/choice information histories. This is followed
by the nonlinear stage where softmax rule is applied to translate the
decision variable into probability for the ensuing Poisson stage. Fitting the model to the data set of each individual allowed us to explore
the individual differences in optimal sensory-value integration in our
task. Our L–N–P model effectively depicted, and generated as well, the
temporal dynamics of human subjects’ perceptual choices made in an
environment with volatile and stochastic feedbacks. The correlation
analysis identified a set of latent model parameters (e.g., reward kernel
weight) that are tightly linked to the individual differences in ability to
adapt their decision to abrupt changes in feedback. The ideal decisionmaker analysis indicated that human subjects are generically suboptimal in forming an effective reward kernel for translating feedbacks in
the past trials into action values in the upcoming trials.
Page 68 of 112
Acknowledgements: Supported by National Research Foundation of
Korea, NRF-2013R1A2A2A03017022.
References
1. Busse L, Ayaz A, Dhruv, NT, Katzner S, Saleem, AB, Schölvinck ML,
Carandini M. The detection of visual contrast in the behaving mouse. J
Neurosci. 2011;31(31):11351–61.
2. Lau B, Glimcher PW. Dynamic response-by-response models of matching
behavior in rhesus monkeys. J Exp Anal Behav. 2005;84(3):555–79.
3. Corrado GS, Sugrue LP, Seung, HS, Newsome, WT. Linear–nonlinear–Poisson models of primate choice dynamics. J Exp Anal Behav. 2005;84(3):581.
P110
A Bayesian algorithm for phoneme Perception and its neural
implementation
Daeseob Lim1, Sang‑Hun Lee1
1
Department of Brain and Cognitive Sciences, Seoul National University,
Seoul, 08826, South Korea
Correspondence: Daeseob Lim ‑ daeseob@snu.ac.kr, visionsl@snu.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P110
Human listeners seem effortless in recognizing a rapid stream of
speech sounds uttered by their fellow speakers, thus being capable
of readily participating in conversation. However, it remains poorly
understood how the brain represents the basic processing unit of such
fluent speech perception, phoneme. In a computational perspective,
phoneme perception is a reverse engineering of speech production,
where the goal is to infer from noisy acoustic signal which phonetic
gesture is the one that was most probably intended by a speaker. This
‘probabilistic inferential’ nature of computations makes the Bayesian
framework attractive. Here we developed a Bayesian algorithm that
captures the two characteristic phenomena of phoneme perception,
(1) sharp transitions in perception (categorization) and (2) enhanced
discriminability (differentiation) at around phoneme boundaries,
and then explored how this algorithm can be implemented in plastic
human brains.
Our model posits (i) that the brain has the probabilistic knowledge
of frequencies of phonetic stimuli prior to forming the likelihood of
phoneme stimuli based on noisy sensory signals (prior and likelihood beliefs), (ii) that the brain combines these two knowledges to
form a posterior distribution of probability (posterior belief ), and (iii)
that the brain ‘optimally’ utilizes this single posterior belief to concurrently perform the categorization and discrimination tasks. The major
latent variables of our interest were the shape of the prior probability distribution function (prior PDF) and the width of the likelihood
PDF. The parameters for these two PDFs were estimated by fitting
the model to the behavioral data in a pair of psychophysical experiments, where human subjects both categorized and discriminated the
acoustic stimuli comprising the cyclic transition among three voiced
stop consonant–vowel syllables, /ba/-/da/-/ga/, varying in the place of
articulation (labial-alveolar-velar). The behaviorally constrained models revealed the prior PDF with the three modes whose peaks correspond to the three prototypical phoneme syllables and the posterior
PDF with large variance.
Having identified the prior and likelihood PDFs used by the optimal
Bayesian listener, we explored plausible neural mechanisms for implementing those PDFs. Going beyond previous attempts, our proposal of
Bayesian implementation offers a formal account for how the unequal
frequency of acoustic stimuli, i.e., stimulus prior, is developmentally
translated into an unequal distribution of sensory neurons via wellknown canonical principles of neural plasticity (‘neural remapping’).
Specifically we propose that sensory neurons, stimulus tuning preferences of which were equally distributed initially, iteratively shift their
tuning curves toward an experienced stimulus (‘attractive shift’) as a
function of their current responsitivity to that stimulus (Fig. 62). The
population distribution of sensory tuning curves that were shaped
by this remapping scenario, when plugged into typical probabilistic
population coding schemes, reproduced qualitatively the human listeners’ performances in the both phoneme tasks. Our model exercise
on tuning width also shed new light on how the optimal tuning width
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 62 Interaction between tuning curve and prior in population
tuning. A Bias map for combinations of tuning curve width and prior
sharpness. Negative bias means that perception of a near-/ba/stimulus was biased toward/ba/, namely categorical perception. Three
inset plots on ordinate and abscissa show the cases of the lowest/
median/highest concentration parameters of tuning curve and prior
peak, respectively. B Discrimination difference map. Negative number
indicates that between-phoneme condition outperformed near-/ba/
condition. C Tuning curves of population neurons that was marked
as green squares in A, B. Location of tuning centers were marked as
dots for 60 neurons, and tuning curves of 30 out of those 60 neurons
were drawn below
of sensory neurons (broad tuning in our case) can be constrained by
the task requirements (categorical perception) and the stimulus environments (biased prior) imposed on a given sensory system (speech
perception).
P111
Complexity of EEG signals is reduced during unconsciousness
induced by ketamine and propofol
Jisung Wang1, Heonsoo Lee1
1
Physics department, Pohang University of Science and Technology,
Pohang, South Korea
Correspondence: Heonsoo Lee ‑ beafool@postech.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P111
Identifying a universal feature of brain dynamics during anestheticinduced unconsciousness has been an important work for both practical use of monitoring depth of general anesthesia and scientific
knowledge about the nature of consciousness. However, it is difficult
because anesthetics with different mechanism of action (MOA) induce
distinct brain dynamics. From the perspective of complex system science, we claim that the dynamics generated by conscious brain is more
complex compared to one from anesthetic-induced unconscious brain
regardless of anesthetic types. To test the hypothesis, we used ketamine and propofol which fall into two distinct anesthetic groups [1].
Disorder and complexity of electroencephalogram (EEG) signals were
analyzed before and after bolus injection of drugs. For the analysis,
we employed Shannon entropy (SE) and fluctuation complexity (FC),
which are information theory-based measures quantifying disorder
and complexity, respectively [2]. The study shows that ketamine and
propofol both reduced the complexity (p < 0.00001 for both) of EEG
signals from the whole brain area (Fp1, F3, T3, P3) while each respectively increased (p = 0.000112) and decreased (p < 0.00001) disorder
of the signal (Fig. 63). The finding supports our claim and suggests considering the EEG complexity as a common measure of consciousness.
Page 69 of 112
Fig. 63 A SE and FC values of Fp1 channel for three different states,
which are wakeful, ketamine-induced, and propofol-induced states,
are averaged over subjects (n = 29 for ketamine-induced, n = 20 for
propofol-induced and n = 49 for wakeful states). Error bars represent
standard errors. Wakeful state has the intermediate SE value between
ones of two other states. For FC value, however, wakeful state has the
highest one and both anesthetized states have smaller ones, forming
a concave relationship between three states. B Each dot manifests
averaged SE and FC values of one subject over 23 10 s-long epochs
overlapping 5 s each other. Error bars here also indicate standard
errors. Most wakeful states have higher FC values compared to ones
of unconscious states and have intermediate SE values. Ketamineinduced states are mainly located in the lower right part when
propofol-induced states are clustered at the lower left part of the
area in SE–FC plot. C FC values of EEG signals from the whole brain
area, covering pre-frontal, frontal, temporal, and parietal regions,
significantly decreased during ketamine-induced loss of consciousness (p < 0.0001). D All FC values of the signals from different regions
were also significantly reduced during propofol-induced unconscious
state (p < 0.0001)
References
1. Lee U, Ku S, Noh G, Baek S, Choi B, Mashour GA. Disruption of frontal–
parietal communication by ketamine, propofol, and sevoflurane. J Am
Soc Anesthesiol. 2013;118:1245–6.
2. Bates JE, Shepard HK. Measuring complexity using information fluctuation. Phys Lett A. 1993;172:416–25.
P112
Self‑organized criticality of neural avalanche in a neural model
on complex networks
Nam Jung1, Le Anh Quang1, Seung Eun Maeng1, Tae Ho Lee1, Jae Woo
Lee1
1
Department of Physics, Inha University, Namgu, Incheon 22212, Korea
Correspondence: Jae Woo Lee ‑ jaewlee@inha.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P112
The concept of the self-organized criticality is applied to many natural and economical systems [1–4]. The distribution of neural avalanche
in neural models obeys a power law with exponents of the mean-field
theory. Neural avalanches in cultured neocortical network show selforganized criticality over long stable period with exponent −1.5 of
the power law for the distribution of the neural avalanche [1]. We consider a modified integrate-and-fire model introduced by Levina, Herrmann and Geisel (LHG model) [2]. We extend the LHG model on the
complex networks such as fully-connected network, random network,
small-world network, and scale-free networks. In the LHG model the
BMC Neurosci 2016, 17(Suppl 1):54
Page 70 of 112
Fig. 65 A Microelectrodes placed to cover both entorhinal cortex
and hippocampus. B Temporal changes in global and local efficiency
around the time of ictal-like epileptiform activity. The red vertical line
indicates the initiation of ictal-like events and the time series in blue
displays recorded field potentials
Fig. 64 Distribution of avalanche size for LHG model on the fullyconnected network. The distribution function of avalanche size
shows the power law with exponent −1.57
membrane potential of a neuron is accumulated from input potential
and random external input. In a fully connected network we observed
the power law with exponent −1.57 as shown in Fig. 64. The exponent of the power law depends on the network structure of the neural
systems.
Acknowledgements: This research was supported by the Basic Science Research Program through the National Research Foundation
of Korea (NRF) funded by the Ministry of Science, ICT and Future Planning (NRF-2014R1A2A1A11051982).
References
1. Beggs J, Plenz D. Neural avalanche in neocortical circuits. J Neurosci.
2003;23:11167–77.
2. Levina A, Herrmann JM, Geisel T. Dynamical synapses causing self-organized criticality in neural networks. Nat Phys. 2007;3:857–60.
3. Li X, Small M. Neuronal avalanches of a self-organized neural network
with active-neuron-dominant structure. Chaos. 2012;22:023104.
4. Liu H, Song Y, Xue F, Li X. Effects of bursting dynamic features on the
generation of multi-clustered structure of neural network with symmetric
spike-timing-dependent plasticity learning rule. Chaos. 2015;25:113108.
P113
Dynamic alterations in connection topology of the hippocampal
network during ictal‑like epileptiform activity in an in vitro rat
model
Chang‑hyun Park1,2, Sora Ahn3, Jangsup Moon1,2, Yun Seo Choi2, Juhee
Kim2, Sang Beom Jun3,4, Seungjun Lee3, Hyang Woon Lee1,2
1
Departments of Neurology, Ewha Womans University School
of Medicine, Seoul, Korea; 2Department of Medical Science, Ewha
Womans University School of Medicine, Seoul, Korea; 3Department
of Electronics Engineering, Ewha Womans University College
of Engineering, Seoul, Korea; 4Brain & Cognitive Sciences, Ewha Womans
University College of Scranton, Seoul, Korea
Correspondence: Chang‑hyun Park ‑ park.changhyun@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P113
The experimental approach using in vitro slices of the rat limbic system has been applied to identify mechanisms underlying epileptiform activity in ictal-like events [1]. We prepared combined slices of
the rat hippocampus-entorhinal cortex and placed them in artificial
corticospinal fluid that contained 4-aminopyridine (4AP). Field potential recordings were made with a microelectrode array composed of
6 × 10 microelectrodes with inter-electrode spacing of 500 μm (see
Fig. 65A) when synchronous activity was induced by 4AP. Channels
with high artifact rates were rejected, and the signal for each remaining channel was divided into 10 s windows without overlap. For each
window, adjacency matrices, or binary networks, were estimated via
inter-channel connections based on spectral coherence in different
frequency bands, including delta, theta, alpha, beta, gamma, ripple,
and fast ripple. Topological properties of the inter-channel networks
were assessed by calculating global and local efficiency [2]. As ictallike events were initiated, global efficiency started to decrease and
local efficiency started to increase, and the changes were maintained
during ictal-like epileptiform activity (see Fig. 65B). Such changes
related to a shift in connection topology to a regularized pattern are in
line with the findings for the whole brain in a rat model [3].
Conclusions Although the initiation and propagation of epileptiform
activity may not be fully appreciated due to the spatially isolated
structure in the in vitro slice preparation, the pattern of ictal-like synchronous activity in the limbic system was related to changes in connection topology that may reflect a shift in brain states.
Acknowledgements: This research was supported by Basic Science
Research Program through the National Research Foundation of
Korea (NRF) funded by the Ministry of Science, ICT & Future Planning
(2015R1C1A1A01052438 to C. Park and 2014R1A2A1A11052103 to H.
W. Lee), and by the Korea Health Technology R&D Project through the
Korea Health Industry Development Institute (KHIDI) funded by the
Ministry of Health & Welfare (HI14C1989 to H. W. Lee).
References
1. Avoli M, Barbarosle M, Lücke A, Nagao T, Lopantsev V, Köhling R: Synchronous GABA-mediated potentials and epileptiform discharges in the rat
limbic system in vitro. J Neurosci. 1996;16(12):3912–24.
2. Latora V, Marchiori M. Efficient behavior of small-world networks. Phys
Rev Lett. 2001;87(19):198701.
3. Otte WM, Dijkhuizen RM, van Meer MPA, van der Hel WS, Verlinde SAMW,
van Nieuwenhuizen O, Viergever MA, Stam CJ, Braun KPJ. Characterization of functional and structural integrity in experimental focal epilepsy:
reduced network efficiency coincides with white matter changes. PLoS
One. 2012;7(7):e39078.
P114
Computational model to replicate seizure suppression effect
by electrical stimulation
Sora Ahn1, Sumin Jo1, Eunji Jun1, Suin Yu1, Hyang Woon Lee2, Sang Beom
Jun1, Seungjun Lee1
1
Department of Electronics Engineering, Ewha Womans University, Seoul,
120‑750, Korea; 2Department of Neurology, Ewha Womans University,
Seoul, 120‑750, Korea
Correspondence: Seungjun Lee ‑ slee@ewha.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P114
Deep brain stimulation (DBS) method for suppression of epileptic seizure is being developed mostly based on clinical experiences because
the suppression mechanism by electrical stimulation is still unclear. As
such, it is difficult to improve efficacy of the DBS method. The study
of computational models allows to predict and analyze the effect of
electrical stimulation by computer simulation such that it can help to
determine optimum stimulation parameters to suppress seizure activity in various conditions.
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 66 Recording data (A) and simulation results (B) of SLE suppression effect by electrical stimulation
In this paper, we propose a hippocampal network model which portrays propagation characteristics of seizure-like events (SLEs) and suppression phenomena by electrical stimulation. The model is composed
of four sub-networks representing EC, DG, CA3 and CA1 and wellknown synaptic pathways between sub-networks. Each sub-network
consists of excitatory and inhibitory neurons which are described by
Izhikevich’s model [1]. Besides synaptic transmission [2], electrical
field transmission [3] between neurons is also considered. Input gains
of neurons are controlled by interaction strengths between sub-networks which are calculated by Granger causality analysis. We adopt
the “potassium accumulation hypothesis” in order to replicate suppression effect by electrical stimulation [4, 5]. The effectiveness of the
model is confirmed by comparing the simulation results with experimental data which were measured in rat hippocampal slice (horizontal, 400um) in bicuculline bath application. Local field potentials are
recorded using micro-electrode array (MEA) and electrical stimulation
(130 Hz, 500 µA, biphasic, 3–5 s) is applied manually in EC by an additional depth electrode when SLE is initiated.
Following Fig. 66 shows time domain signals recorded in in vitro measurement and generated from the computer model, respectively. After
stimulation, the SLE in EC is suppressed immediately, while SLEs in
other areas still remained. The simulation results show similar waveforms with experimental data.
Acknowledgements: This work was supported by the National
Research Foundation of Korea (No. 2014R1A2A1A11052763).
References
1. Izhikevich EM. Simple model of spiking neurons. IEEE Trans Neural Netw.
2003;14(6): 1569–72.
2. Izhikevich EM, Gally JA, Edelman GM. Spike-timing dynamics of neuronal
groups. Cereb Cortex. 2004;14(8):933–44.
3. Fröhlich F, McCormick DA. Endogenous electric fields may guide neocortical network activity. Neuron. 2010;67(1):129–43.
4. Fertziger AP, Ranck JB. Potassium accumulation in interstitial space during
epileptiform seizures. Exp Neurol. 1970;26(3):571–85.
5. Beurrier C, Bioulac B, Audin J, Hammond C. High-frequency stimulation
produces a transient blockade of voltage-gated currents in subthalamic
neurons. J Neurophysiol. 2001;85(4):1351–56.
P115
Identifying excitatory and inhibitory synapses in neuronal
networks from spike trains using sorted local transfer entropy
Felix Goetze1,2, Pik‑Yin Lai1
1
Department of Physics, National Central University, Chung‑Li, Taiwan,
ROC; 2Taiwan International Graduate Program for Molecular Science
and Technology, Institute for Atomic and Molecular Sciences, Academia
Sinica, Taipei, Taiwan, ROC
Correspondence: Felix Goetze ‑ afgoetze@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P115
Transfer entropy [1] is the established method for quantifying the
effective connectivity among neurons.
It has been shown in simulations [2] that measuring it from simultaneously recorded spike trains of neurons can detect the underlying
connections and therefore reconstruct a neuronal network from its
observed dynamics. Being interpreted as the predicted information
Page 71 of 112
transfer, it quantifies the directed non-linear interactions between
time series as a model-free method regardless of the underlying interaction type, which could be either inhibitory or excitatory.
Making the distinction between excitatory and inhibitory synapses,
however is important in order to understand the underlying principles
of spatiotemporal patterns in functional networks.
In our study we describe a method for the measuring of interaction
types, based on the concept of local transfer entropies [3]. In contrast
to the averaging across all configurations of variables of the source
process and the target process as in the Transfer Entropy estimation,
the local transfer entropy quantifies the effect of a specific configuration of variables on how they either inform or misinform on the future
of the target process. For example observing the presynaptic neuron
fire and then observing the postsynaptic neuron fire is informative for
an excitatory connection, but misinformative for an inhibitory connection. On average, knowing the past of the source process of an inhibitory or excitatory connection are both predictive of the future of the
target process, but local transfer entropies of specific variable configurations have opposite signed values for each interaction type respectively. Sorting the local entropies according to interaction type yields
the quantity we call sorted local transfer entropy that aims to identify
inhibitory and excitatory synapses from recorded spike trains.
We validate this method with simulated spike trains from the Izhikevich
model [4] of cortical neuronal networks, by following a previous paper
[2]. The random network is noise-driven and consists of 800 excitatory and 200 inhibitory neurons. Synapses have random delays and
the synaptic strengths evolved according to a spike-timing plasticity
rule, before the recordings for the analysis are collected. Using Transfer
Entropy and the new quantity Sorted Local Transfer Entropy, we reconstruct the networks and distinguish inhibitory from excitatory synapses.
The use of two decision boundaries for classifying inhibitory and excitatory synapses separately improves the overall network reconstruction.
References
1. Schreiber T. Measuring information transfer. Phys Rev Lett. 2000;85:461.
2. Ito S, Hansen ME, Heiland R, Lumsdaine A, Litke AM, Beggs JM, Zochowski
M. Extending transfer entropy improves identification of effective connectivity in a spiking cortical network model. PLoS One. 2011;6:e27431
3. Lizier JT. Measuring the dynamics of information processing on a local
scale in time and space. In: Directed information measures in neuroscience. Berlin: Springer; 2014. p. 161–93.
4. Izhikevich EM. Polychronization: computation with spikes. Neural Comput. 2006;18(2):245–82.
P116
Neural network model for obstacle avoidance based
on neuromorphic computational model of boundary vector cell
and head direction cell
Seonghyun Kim1, Jeehyun Kwag1
Department of Brain and Cognitive Engineering, Korea University, Seoul,
Korea
Correspondence: Jeehyun Kwag ‑ jkwag@korea.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P116
Developing robots that can perform autonomous exploration in unfamiliar environment is one of the challenges in robotics. Autonomously
navigating robots are generally equipped with an obstacle avoidance
(OA) system based on sensors such as Light Detection and Ranging
(LIDAR) and camera to detect obstacles as well as complex algorithms
to correct the noise of sensors [1]. Interestingly, rodent brain shows
remarkable reliability to noise in OA [2] through using neurons specialized in processing spatial orientation and spatial boundary called head
direction cell (HDC) [3] and boundary vector cell (BVC) [4], respectively.
Therefore, building a bioinspired OA system with neural network that
consists of neuromorphic HDC and BVC may help increase the efficiency of autonomous navigation. Hence, we built a neural network for
OA consisting of all-to-all synaptic connections between six HDCs, four
BVCs, and two motor neurons where HDC and BVC were constructed
as multi-compartment Hodgkin–Huxley models using the NEURON
based on full morphology and electrophysiological properties in vitro
BMC Neurosci 2016, 17(Suppl 1):54
[5, 6]. Each HDC was modeled to spike at specific preferred directions
separated by 60°, BVCs were modeled to spike at boundaries of cardinal directions and motor neurons were modeled to spike with Gaussian white noise as a background noise. We also built a virtual rat that
navigated within a 1 m × 1 m environment whose trajectory was
controlled by spikes of motor neurons receiving synaptic inputs from:
(1) HDCs, (2) BVCs and (3) both HDCs and BVCs. Number of obstacle
detection (detection number: DN) and the time spent during obstacle
collision (collision time: CT) were analyzed to compare the efficiency
of neural network for OA.
We first verified that our neuromorphic HDC and BVC models could
mimic the experimentally recorded electrophysiological properties
in vitro and in vivo: HDCs reached maximum firing rate at each preferred direction, and BVCs increased their firing rate as the virtual rat
approached boundaries. Using such neuromorphic HDC and BVC models, we investigated the roles of HDC and BVC in neural network for OA.
Firstly, we performed the control simulation where virtual rat was controlled with neural network without neither HDC nor BVC and observed
that DN was 40 and CT was 137 s. When HDC was added to the neural network, the result was similar to control simulation, (DN = 38
and CT = 139 s), indicating HDC alone cannot perform OA efficiently.
When BVC alone was included in the neural network, DN substantially
increased and CT decreased compared to control model (DN = 110
and CT = 73 s), indicating that OA efficiency increased. Finally, when
both HDC and BVC were included in the neural network, the OA performance was most efficient (DN = 139 and CT = 39 s). These results suggest that our neural network model composed of neuromorphic HDC
and BVC neurons can successfully perform OA even with background
noise. Therefore, here we suggest the bioinspired neural network that
consists of neuromorphic computational model of HDC and BVC could
serve as a new approach to build an efficient OA system.
Acknowledgements: This study was supported by the Basic Science Research Program through the National Research Foundation
of Korea funded by the Ministry of Science, ICT and Future Planning
(NRF-2013R1A1A2053280).
References
1. Zohaib M, Pasha M, Riaz R, Javaid N, Ilahi M, Khan R. Control strategies for mobile robot with obstacle avoidance. J Basic Appl Sci Res.
2013;3(4):1027–36.
2. Vorhees CV, Williams MT. Assessing spatial learning and memory in
rodents. ILAR J. 2014;55(2):310–32.
3. Taube JS, Muller RU, Ranck JB Jr. Head-direction cells recorded from the
postsubiculum in freely moving rats. I. Description and quantitative
analysis. J Neurosci. 1990;10(2):420–35.
4. Lever C, Burton S, Jeewajee A, O’Keefe J, Burgess N: Boundary vector cells in the subiculum of the hippocampal formation. J Neurosci.
2009;29(31):9771–77.
5. Yoder RM, Taube JS. Projections to the anterodorsal thalamus and lateral
mammillary nuclei arise from different cell populations within the
postsubiculum: implications for the control of head direction cells. Hippocampus. 2011;21(10):1062–73.
6. Menendez de la Prida L, Suarez F, Pozo MA. Electrophysiological and morphological diversity of neurons from the rat subicular complex in vitro.
Hippocampus. 2003;13(6):728–44.
P117
Dynamic gating of spike pattern propagation by Hebbian
and anti‑Hebbian spike timing‑dependent plasticity in excitatory
feedforward network model
Hyun Jae Jang1, Jeehyun Kwag1
Department of Brain and Cognitive Engineering, Korea University, Seoul,
Korea
Correspondence: Jeehyun Kwag ‑ jkwag@korea.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P117
Precise timings of spikes within in vivo spike train are believed to
carry information critical for neural computation [1]. For such neural
Page 72 of 112
information to be effective, temporal patterns of spike train should
be able to propagate across multiple neuronal layers in the feedforward network (FFN) of the brain without dissipation [2]. To support
such reliable propagation of spike patterns, preferential and selective
strengthening of synaptic pathways through which the spike patterns
are routed may be necessary. Asymmetric Hebbian and anti-Hebbian
spike timing-dependent plasticity (STDP), where synaptic strengths
are strengthened or weakened depending on the precise relative timing and the order between pre- and postsynaptic spikes [3, 4], may
serve as a good candidate for dynamically routing the propagation of
spike patterns.
Hence, we investigated the role of Hebbian and anti-Hebbian STDP in
spike pattern propagation using a six-layered FFN model composed
of 200 Hodgkin–Huxley excitatory neurons in each layer. Asymmetric and symmetric/Hebbian and anti-Hebbian STDP were modeled at
excitatory synapses using exponential functions [5]. In vivo spike train
obtained from public database (crcns.org) was used as an input spike
pattern (TIN) in layer 1, which was simulated in a small subset of excitatory neurons in layer 1 of FFN model, while the rest were made to
spontaneously spike with spike frequencies showing log-normal distribution to mimic in vivo background noise. The propagation of temporal spike pattern was quantified by analyzing the similarity ratio (SR)
between TIN and output spike pattern in layer 6 (TOUT ), which calculates how instantaneous inter-spike intervals of TIN and TOUT are similar.
In FFN model without STDP, the spike pattern of TIN in layer 1 became
dissipated in noise as it propagated across layers, and consequently
failed to preserve its spike pattern to layer 6 with low SR (0.49). When
asymmetric anti-Hebbian STDP was included in FFN model, TIN also
failed to propagate to layer 6 with low SR (0.17). However, in the presence of asymmetric Hebbian STDP, TIN successfully propagated to the
final layer with high SR (0.87), indicating that asymmetric Hebbian
STDP preferentially enhanced TIN propagation in FFN model. Further
analysis revealed that asymmetric Hebbian STDP selectively strengthened the synaptic weights of the synaptic pathways routing TIN while
it weakened the synaptic weights of that routing noise, effectively
serving as an open-gate for propagating TIN. In contrast, asymmetric
anti-Hebbian STDP curve selectively weakened the synaptic weights
of the synaptic pathways routing TIN, serving as a close-gate for propagating TIN. We also tested the effect of symmetric Hebbian STDP which
induces only LTP or symmetric anti-Hebbian STDP which induces only
LTD, and found that both types of symmetric STDP failed to propagate TIN with low SR (symmetric Hebbian = 0.23, symmetric anti-Hebbian = 0.14). Our results demonstrate that only asymmetric Hebbian
STDP facilitates the reliable propagation of in vivo temporal pattern
while asymmetric and symmetric anti-Hebbian STDP blocks temporal pattern propagation, suggesting that different types of STDP may
dynamically gate the propagation of neural information.
Acknowledgements: This study was supported by Human Frontier Science Program (RGY0073/2015) and the Basic Science
Research Program through the National Research Foundation of
Korea funded by the Ministry of Science, ICT and Future Planning
(NRF-2013R1A1A2053280).
References
1. Mainen ZF, Sejnowski TJ. Reliability of spike timing in neocortical neurons.
Science. 1995;268(5216):1503–6.
2. Kumar A, Rotter S, Aertsen A. Spiking activity propagation in neuronal
networks: reconciling different perspectives on neural coding. Nat Rev
Neurosci. 2010;11(9):615–27.
3. Bi GQ, Poo MM. Synaptic modifications in cultured hippocampal neurons:
dependence on spike timing, synaptic strength, and postsynaptic cell
type. J Neurosci. 1998;18(24):10464–72.
4. Feldman DE. The spike-timing dependence of plasticity. Neuron.
2012;75(4):556–71.
5. Song S, Miller KD, Abbott LF. Competitive Hebbian learning
through spike-timing-dependent synaptic plasticity. Nat Neurosci.
2000;3(9):919–26.
BMC Neurosci 2016, 17(Suppl 1):54
P118
Inferring characteristics of input correlations of cells exhibiting
up‑down state transitions in the rat striatum
Marko Filipović1,2, Ramon Reig3, Ad Aertsen1,2, Gilad Silberberg4, Arvind
Kumar1,5
1
Bernstein Center Freiburg, Freiburg, Germany; 2Faculty of Biology,
University of Freiburg, Freiburg, 79104, Germany; 3Instituto de
Neurociencias de Alicante, University of Alicante, Alicante, Spain;
4
Department of Neuroscience, Karolinska Institute, Stockholm, 17177,
Sweden; 5Department of Computational Science and Technology,
School of Computer Science and Communication, KTH Royal Institute
of Technology, Stockholm, 10040, Sweden
Correspondence: Marko Filipović ‑ marko.filipovic@bcf.uni‑freiburg
BMC Neuroscience 2016, 17(Suppl 1):P118
Rat striatal projection neurons (SPNs) recorded under ketamine anesthesia exhibit slow oscillations, with transitions between depolarized and hyperpolarized membrane potential also referred to as up
and down states, respectively. It is presumed that the activity during
hyperpolarized down-states is determined by intracellular processes,
whereas the large membrane voltage fluctuations during up-states
are a product of increased synaptic input. Because local striatal activity
during an up-state is weak, the statistics of the up-state fluctuations
mainly reflect cortical feedforward input to the SPNs.
To infer the statistics of the cortical input to SPNs we measured the
statistics and spectrum of the membrane potential of SPNs in up and
down states. The spectrum of the membrane potential reflects the
filtering properties of the membrane and can be used to estimate
the effective time constant (τeff ) of the neuron. Our analysis showed
that SPNs have significantly smaller τeff in the up-state than in the
down-state, consistent with the assumption that the barrage of synaptic input causes an increase in membrane conductance during the
up state. However, this observation is inconsistent with the idea that
depolarization of SPNs should increase the membrane time constant
because of the closing of some of the voltage dependent ion channels
(e.g. the Kir) channels [1].
The mean (μup) and variance (σup) of the membrane potential during
up states varied in a correlated manner. At the same time, for a given
SPN, μup and σup of individual up-state membrane potentials were
highly variable across different up states, indicating a corresponding
variability in the cortical inputs. Using a point neuron model of an SPN,
we show that the correlation and variability of the up-state mean and
variance could be explained if we assume that SPNs receive correlated
inputs.
Across different SPNs, each recorded in a different animal, we observed
a high variability in the correlation (ρ) between μup and σup. This variability could arise from the heterogeneity in the neuron morphology,
intracellular properties, conductance state of the neurons, synaptic
weights and the input rate and correlations. Using a point neuron
model we tested the dependence of ρ on each of these properties.
Our analysis showed that the variability of the correlation between μup
and σup arises because of the diversity of synaptic weights and input
correlations, and not because of intrinsic properties of SPNs. This suggests that neuronal heterogeneity could be obscured by the statistics
of the synaptic inputs and synaptic weights.
In summary, our analysis of up-down states allows us to make general
inferences about characteristics of correlated synaptic input, such as
strength of correlations and input firing rate, solely based on membrane potential recordings of SPNs exhibiting up and down states.
Acknowledgements: This work was supported in parts by the Erasmus Mundus Joint Doctorate Programme EUROSPIN (MF), an ERC
starting grant (GS), the Knut and Alice Wallenberg Academy Fellowship, the Karolinska Institutet Strategic Research program in Neuroscience (StratNeuro; GS, AK), and the Swedish Medical Research Council
(GS).
Reference
1. Nisenbaum ES, Wilson CJ. Potassium currents responsible for inward
and outward rectification in rat neostriatal spiny projection neurons. J
Neurosci. 1995, 15: 4449–63.
Page 73 of 112
P119
Graph properties of the functional connected brain under the
influence of Alzheimer’s disease
Claudia Bachmann1, Simone Buttler1, Heidi Jacobs2,3,4, Kim Dillen5, Gereon
R Fink5,6, Juraj Kukolja5,6, Abigail Morrison1,7,8
1
Institute of Neuroscience and Medicine (INM‑6) and Institute
for Advanced Simulation (IAS‑6) and JARA BRAIN Institute I, Jülich
Research Centre, Jülich, Germany; 2Faculty of Health, Medicine and Life
Science, School for Mental Health and Neuroscience (MHeNS), Alzheimer
Centre Limburg, Maastricht University Medical Centre, PO Box 616, 6200
MD Maastricht, The Netherlands; 3Department of Radiology & Athinoula
A. Martinos Center for Biomedical Imaging, Massachusetts General
Hospital, Harvard Medical School, Boston, MA 02114, USA; 4Faculty
of Psychology and Neuroscience, Department of Cognitive Neuroscience,
Maastricht University, PO BOX 616, 6200 MD Maastricht, The Netherlands;
5
Cognitive Neuroscience, Institute of Neuroscience and Medicine (INM‑3),
Jülich Research Centre, Jülich, Germany; 6Department of Neurology,
University Hospital of Cologne, Cologne, Germany; 7Computational
Neuroscience, Bernstein Center Freiburg, Freiburg, 79104, Germany;
8
Institute of Cognitive Neuroscience, Faculty of Psychology,
Ruhr‑University Bochum, 44801 Bochum, Germany
Correspondence: Claudia Bachmann ‑ c.bachmann@fz‑juelich.de
BMC Neuroscience 2016, 17(Suppl 1):P119
Diagnosing Alzheimer’s disease (AD), especially in the early stage, is
costly and burdensome for the patients, since it comprises a battery
of psychological tests and an extraction of disease specific biomarkers
from the cerebrospinal fluid. A cheaper and more convenient procedure would be a diagnosis based on images obtained through fMRI.
Based on previous polymodal studies demonstrating disrupted interand intra-cortical connectivity in AD [1], we argue that the functional
connectivity of the whole cortex might be a good predictor for the
cause of the disease. In resting state fMRI, previous attempts to analyze graph properties of whole brain networks contradict each other
[2]. In our opinion there are two general critical points in the methodology of these studies that are likely to contribute to the variability of the results. First, we criticize that the activities of the brain areas
(graph nodes) that are used to calculate the functional connectivities
(weights of the graph edges) are composed of functionally inhomogeneous signals, as individual brains are often mapped onto a standard
atlas brain of known functional coherent areas [2, 3]. The second problem consists in converting the resulting weighted graphs into simple
graphs, by setting weights above an arbitrary threshold wmin to 1, and
those below it to 0 [2]. The drawback here is that there is no validation
for an optimal threshold, and information that might be relevant in
AD may be lost. In this work we address the first problem by applying
an activity-driven, region-growing clustering algorithm derived from
image processing [4]. In order to guarantee functionally homogeneous clusters, the threshold for inclusion of a voxel in a region is regulated by a heterogeneity criterion [3]. Applying this algorithm, we end
up with undirected weighted graphs with varying numbers of nodes
for three sets of data: healthy elderly controls, mild cognitive impairment and Alzheimer’s disease. Targeting the second problem, we
analyze the dependence of graph theoretic measures (shortest path
length, in- and out-degree distribution, clustering coefficient, modularity and minimal spanning tree [5]) on wmin. Finally, we investigate
the distribution of these measures for each data set to determine candidates for a predictive measure.
Acknowledgements: We acknowledge partial support by the Helmholtz Alliance through the Initiative and Networking Fund of the
Helmholtz Association and the Helmholtz Portfolio theme “Supercomputing and Modeling for the Human Brain”.
References
1. Bokde AL, Ewers M, Hampel H.: Assessing neuronal networks: understanding Alzheimer’s disease. Prog Neurobiol. 2009;89:125–33.
2. Tijms BM, Wink AM, de Haan W, van der Flier WM, Stam CJ, Scheltens
P, Barkhof F. Neurobiol: Alzheimer’s disease: connecting findings from
graph theoretical studies of brain networks. Aging. 2013;34: 2023–36.
BMC Neurosci 2016, 17(Suppl 1):54
3.
4.
5.
Marrelec G, Fransson P. Assessing the influence of different ROI selection
strategies on functional connectivity analyses of fMRI data acquired during steady-state conditions. PLoS One. 2011;6(4):e14788.
Lu Y, Jiang T, Zang Y. Region growing method for the analysis of functional MRI data. Neuroimage. 2003;20(1):455–65.
Wang J, Zuo X, Dai Z, Xia M, Zhao Z, Zhao X, Jia J, Han Y, He Y. Disrupted
functional brain connectome in individuals at risk for Alzheimer’s disease.
Biol Psychiatry. 2013;73(5):472–81.
P120
Learning sparse representations in the olfactory bulb
Daniel Kepple1, Hamza Giaffar1, Dima Rinberg2, Steven Shea1, Alex
Koulakov1
1
Cold Spring Harbor Laboratory, Cold Spring Harbor, NY 11724, USA;
2
NYU Neuroscience Institute, New York, NY 10016, USA
Correspondence: Daniel Kepple ‑ akula@cshl.edu
BMC Neuroscience 2016, 17(Suppl 1):P120
In mouse olfaction, olfactory receptor responses are aggregated in
spherical structures of the main olfactory bulb (MOB) called ‘glomeruli.’
These signals are then received by mitral cells (MC) and communicated
to the cortex. MC responses are modified by local inhibitory interneurons called granule cells GC, which vastly outnumbered MCs (50fold–
100fold). In our previous work [1], we proposed a model in which GCs
inhibit MC to form a sparse, incomplete representation (SIR) of odors
(Fig. 67). Here, we reason that since sparse representations are efficient, sparseness may increase with learning. We extend the SIR model
to allow network synaptic weights to be adjusted, increasing representational sparseness and increasing stimulus discriminability. We derive
learning rules for dendrodendritic connectivity between GCs and MCs
and also for centrifugal cortico-granule synapses. We computationally
test these learning rules and make several predictions of GC and MC
plasticity. Specifically, we predict that a minority of GCs outcompete
the rest of the population to generate a negative image of a learned
odor. Additionally, we predict that participation of the GC network will
confer combination selectivity and the ability to discriminate overlapping input patterns. Finally, we experimentally validate these predictions for the dynamics of GCs during locus coeruleus-induced MOB
plasticity.
Reference
1. Koulakov AA, Rinberg D. Sparse incomplete representations: a potential
role of olfactory granule cells. Neuron. 2011;72(1):124–36
Fig. 67 Sparse incomplete representations (SIR). In our previously
formulated model of the main olfactory bulb network [1], MCs
receive inputs from receptor neurons in the glomeruli (black circles)
and interact with GCs through dendrodendritic synapses. GCs build
representations of MC glomerular inputs (red arrows). The representations are contained in the inhibitory inputs returned by the GCs
to the MCs (blue arrows). Because GCs inhibit each other through
second-order inhibitory interactions, only a few GCs respond to an
odorant (full blue circles with a dendrite shown). The vast majority of
GCs do not change their firing rate in response to an odorant (empty
circles). Thus, the responses of GCs are sparse. Because some MCs
manage to retain the responses to odorants, the representation by
GCs is called incomplete. According to this model, MCs transmit to
higher areas the errors in the GC representation
Page 74 of 112
P121
Functional classification of homologous basal‑ganglia networks
Jyotika Bahuguna1,2,3, Tom Tetzlaff1, Abigail Morrison1,2, Arvind Kumar2,3,
Jeanette Hellgren Kotaleski3
1
Institute of Neuroscience and Medicine (INM‑6), Institute for Advanced
Simulation (IAS‑6) and JARA BRAIN Institute I, Jülich Research Centre,
Jülich, Germany; 2Computational Neuroscience, Bernstein Center
Freiburg, Freiburg, 79104, Germany; 3Computational Brain Science,
Department of Computational Science and Technology, School
of Computer Science and Communication, KTH, Royal Institute
of Technology, Stockholm, Sweden
Correspondence: Jyotika Bahuguna ‑ j.bahuguna@fz‑juelich.de
BMC Neuroscience 2016, 17(Suppl 1):P121
The basal ganglia (BG) are a set of nuclei that play an important role
in motor and cognitive functions. Indeed many brain diseases such
as Parkinson’s disease (PD) can be attributed to dysfunction of one
or more BG nuclei. The classical model of basal ganglia has been
regularly updated with discoveries of new sub-populations within a
nucleus or new projections from existing nuclei in recent years. It is
unclear how these new insights on the structure of the BG network
foster our understanding of its function. The effective connectivities
among these recently identified BG sub-populations are only partially
known. In the framework of a simple firing-rate model subjected to a
genetic algorithm, we identified effective BG connectivities which are
consistent with experimentally established firing-rate and phase relationships in Subthalamic Nucleus (STN) and two GPe subpopulations
(arkypallidal [GPe-TA] and prototypical [GPe -TI]) in both healthy and
PD states [1]. This is in extension to an earlier model that identified
effective connectivities for the STN-TA-TI-sub circuit [2].
As expected, we found that multiple parameter combinations can
fit the data [1]. We re-classified these homologous networks that
reproduced the healthy and PD state, on the basis of two dynamical features: suppression of GPi activity and susceptibility of the BG
network to oscillate in the presence of cortical input. These features
were chosen because task execution requires GPi suppression while
oscillations in the STN-GPe subnetwork are characteristic of PD.
We found that most putative pathological networks showed insufficient suppression of GPi activity and high susceptibility to oscillations whereas most putative healthy networks showed sufficient
suppression of GPi activity and low susceptibility to oscillations.
This is consistent with experimental data that shows that lack of GPi
suppression [3] or oscillations [4, 5] is correlated with Parkinsonian
symptoms such as stymied movement and tremor. A small fraction
of networks, however, in both cases show deficiency in only one of
the features. This could indicate the configurations of healthy networks that might be more pathology prone and in contrast configurations of pathological networks that might be easier to push into a
healthy state. Further analysis of estimated BG connectivity revealed
that transitions between the putative PD and healthy networks were
possible by modifying the strength of the relevant projections. Most
of the transitions involved changes in corticostriatal, striatopallidal
and pallidopallidal projections. Finally, the variance observed in the
functional classification of putative pathological and healthy networks might hint at the variance observed in manifestation of Parkinson’s disease (PD).
Acknowledgements: Klinische Forschergruppe (KFO219, TP12) of the
Deutsche Forschungsgemeinschaft; Helmholtz Association, EuroSPIN
and Erasmus Mundus Joint Doctorate Programme.
References
1. Mallet A, Pogosyan A, Márton LF, Bolam JP, Brown P, Magill PJ. Parkinsonian beta oscillations in the external globus pallidus and their relationship with subthalamic nucleus activity. J Neurosci. 2008;28(52):14245–58.
2. Nevado-Holgado AJ, Mallet N, Magill PJ, Bogacz R. Effective connectivity
of the subthalamic nucleus-globus pallidus network during Parkinsonian
oscillations. J. Physiol. 2014;592(7):1429–55.
3. Boraud T, Bezard E, Bioulac B, Gross CE. Ratio of inhibited-to-activated
pallidal neurons decreases dramatically during passive limb movement in
the MPTP-treated monkey. J Physiol. 2000;83(3):1760–63.
BMC Neurosci 2016, 17(Suppl 1):54
4.
5.
Page 75 of 112
Chen CC, Litvak V, Gilbertson T, Kühn A, Lu CS, Lee ST, Tsai CH, Tisch S,
Limousin P, Hariz M, Brown P. Excessive synchronization of basal ganglia
neurons at 20 Hz slows movement in Parkinson’s disease. Exp Neurol.
2007;205(1):214–21.
Moran A, Bergman H, Israel Z, Bar-Gad I. Subthalamic nucleus functional
organization revealed by parkinsonian neuronal oscillations and synchrony. Brain. 2008;131(Pt-12):3395–409.
P122
Short term memory based on multistability
Tim Kunze1,2, Andre Peterson3, Thomas Knösche1
1
Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig,
Germany; 2Institute of Biomedical Engineering and Informatics, Ilmenau
University of Technology, Ilmenau, Germany; 3Department of Medicine,
University of Melbourne, Melbourne, Australia
Correspondence: Tim Kunze ‑ tkunze@cbs.mpg.de
BMC Neuroscience 2016, 17(Suppl 1):P122
Neural circuits can be formally described by modeling the collective
behavior of relatively homogeneous neural populations, so-called
neural masses [1]. Some neural mass models consider the minimum set of one excitatory and one inhibitory subpopulation each.
In contrast, three-population models make a distinction between
excitatory pyramidal cells (PC), projecting to distant areas, and excitatory interneurons (EIN), providing local feedback. We investigate
a three-population neural mass model, driven by input to the EIN,
with respect to its input/output behavior. We find that such a circuit
exhibits, for sufficiently salient inputs, a memory effect based on
multi-stability. Furthermore, we test the hypothesis that this mechanism essentially depends on the separation between input and output neurons and is thus not captured in the simpler two-population
model.
We use a neural mass model [2], where a pyramidal cell subpopulation
receives negative feedback from an inhibitory interneuron subpopulation and positive feedback either directly through self-connections
or indirectly via a secondary excitatory subpopulation of interneurons. The respective feedback topology of interest (including which
subpopulation is targeted by external input) is controlled by a single
parameter. We systematically applied transient sensory inputs, modeled by pulses of various magnitude and duration, as external inputs
to the EIN and monitored the behavior of the PC.
Depending on the duration and intensity of the applied stimuli (see
Fig. 68A), the output either transiently follows the input (i) or it jumps
to a more depolarized state, where it remains oscillating with a higher
mean membrane potential even after the stimulus has ceased (ii) and
where further input does not effect the output any more (iii). This state
can be terminated by an impulse to the inhibitory interneurons (iv).
The accessibility of this memory effect depends on the saliency of the
stimulus in terms of duration and intensity (see Fig. 68B) and disappears in case of direct feedback in a structurally similar two population
model.
The identified short-term memory mechanism would be important
for temporal integration in cortical processing, potentially applicable in predictive coding schemes. The distinction between the
input receiving excitatory subpopulation and the output sending
excitatory subpopulation appears to be crucial for the described
mechanism, which is further modulated by inhibitory feedback.
The further examination of the ratio between excitation and inhibition, governing this mechanism, thus represents an important step
to elucidate how the topology between excitatory and inhibitory
neural populations affects emerging dynamics on a mesoscopic
scale with potential effects on brain states and higher-order brain
functionality.
References
1. Freeman WJ. Mass action in the nervous system. New York: Academic
Press; 1975.
2. Spiegler A, Kiebel SJ, Atay FM, Knösche TR. Bifurcation analysis of neural
mass models: Impact of extrinsic inputs and dendritic time constants.
NeuroImage. 2010;52:1041–58.
Fig. 68 A Response of pyramidal cells to transient input to excitatory
interneurons shows different modes. B Depending on intensity and
duration of the stimulus
P123
A physiologically plausible, computationally efficient model
and simulation software for mammalian motor units
Minjung Kim1, Hojeong Kim1
1
Division of IoT and Robotics Convergence Research, DGIST, Daegu,
42988, Korea
Correspondence: Hojeong Kim ‑ hojeong.kim03@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P123
Background A spinal motoneuron contacts a bunch of muscle fibers
forming a motor unit that underlies all mammalian movements. The
essential role of the motor unit is the transduction of synaptic inputs
from descending and reflex pathways into muscle force. Since the
input–output properties of both motoneurons and muscle fibers are
non-linear, it has been difficult to make predictions on how changes in
synaptic inputs to motoneuron, cellular properties of the motoneuron
and muscle fibers and muscle length may affect motor output [1].
Methods To tackle this fundamental issue in the field of motor neuroscience, we developed a physiologically plausible but computationally efficient model of the motor unit and a software package that
allows for virtual experiments on the input–output properties of the
motor unit over a full range of physiological inputs and biophysical
parameters.
Results The computational model of motor unit was first built in this
study coupling the motoneuron model and the muscle unit model
with a simplified axon model. The motoneuron model was developed
using the recently reported two-compartment modeling approach
[2]. The key feature of the new reduced motoneuron model is that all
cable parameters of the reduced model are analytically determined
based on the system properties such as input resistance, membrane
time constant and electrical coupling properties between the soma
and the dendrites, which are all empirically measurable from real
motoneurons.
For the muscle unit, the recently developed muscle modeling
approach was employed that consists of three sub-modules representing [3]: (1) the transformation of the spike signals from motoneurons
BMC Neurosci 2016, 17(Suppl 1):54
into the dynamics of calcium concentration in the sarcoplasm, (2) the
conversion of the calcium concentration to the muscle activation level,
and (3) the transformation of the muscle activation level into the muscle force using Hill-type muscle mechanics. The new muscle model
was constructed in this study to reflect all experimentally identified
dependencies of muscle activation dynamics on muscle length and
movement over a full range of stimulation frequencies in cat soleus
muscles.
Then, to enhance the usability and extendibility the software package for simulating and analyzing the developed motor unit model was
designed and implemented based on the object-oriented programing paradigm and open source Python language along with graphic
user interfaces (GUI). The software package developed in this study
provides a GUI-based simulation environment in which a single motoneuron, muscle unit, and motor unit can be individually simulated and
analyzed in a wide range of experimental conditions reflecting biological realisms.
Conclusions Our model of the motor unit and user-friendly simulation
software may provide not only a computational framework to gain
systemic insights into motor control by the central nervous system in a
cellular perspective but also a basis on which to build biologically realistic large-scale neuro-musculo-skeletal models.
Acknowledgements: This work was supported by the DGIST R&D Program of the Ministry of Science, ICT and Future Planning of Korea (15RS-02 and 16-RS-02).
References
1. Heckman CJ, Enoka RM: Motor unit. Compr Physiol. 2012;2(4):2629–82.
2. Kim H, Jones KE, Heckman CJ. Asymmetry in signal propagation between
the soma and dendrites plays a key role in determining dendritic excitability in motoneurons. PLoS One. 2014;9(8):e95454.
3. Kim H, Sandercock TG, Heckman CJ. An action potential-driven model
of soleus muscle activation dynamics for locomotor-like movements. J
Neural Eng. 2015;12(4).
P125
Decoding laser‑induced somatosensory information from EEG
Ji Sung Park1, Ji Won Yeon, Sung‑Phil Kim1
1
Department of Human Factors Engineering, Ulsan National Institute
of Science and Technology, Ulsan 689‑798, South Korea
Correspondence: Sung‑Phil Kim ‑ spkim@unist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P125
Recently, our research group has proposed a new way of providing a
non-nociceptive tactile sensation with laser [1]. In this study, we aimed
to investigate laser-induced somatosensory information represented
in cortical activity using the human EEG. The EEG data were acquired
using the V-Amp amplifier (Brain Products GmbH, Gilching, Germany)
with 16 wet electrodes that were placed on the scalp following the
international 10–20 system. Twenty one subjects participated in the
study (7 female and mean age of 22.4 years). During the experiment,
a mechanical stimulus, a laser stimulus and a heat stimulus were
given in a random order to subjects sixty times per stimulus. Subjects described the feeling of laser stimulation as non-painful sensation, painful sensation and no sensation. As described in the previous
study, 56.3, 12.3 and 31.4 % of the subjects reported laser stimulation
as non-painful, painful and no sensation, respectively [1]. To examine similarity of cortical activity in response to different stimuli, we
employed a decoding analysis of the EEG data. In the decoding analysis, we used the linear discriminant analysis (LDA) method to classify
the beta (21–28 Hz) event-related desynchronization/synchronization
(ERD/S) patterns of EEG into one of the two classes representing every
pair of stimuli (a total of six pairs from four stimuli) [2]. Classification
error indicated how similar beta ERD/S patterns were between two
stimuli: a larger error reflected more difficulty in discriminating patterns and consequently a greater similarity between patterns. The beta
ERD/S patterns were estimated using the short time Fourier transform.
Baseline correction was implemented using the 0.5 s period before
stimulus onset. For each pair of stimuli, one-way ANOVA was used to
Page 76 of 112
Fig. 69 Each node represents stimulation type and each edge means
classification error rate. Length of edge shows similarity between a
pair of stimulation
select four channels that exhibited the most differences in beta ERD/S
patterns between classes and classification accuracy was assessed by
the leave-four-out cross validation [3] (see Fig. 69 for the classification
error between every stimulus pair). The classification results showed
that to the beta ERD/S pattern induced by mechanical stimulation,
the pattern by non-painful laser stimulation was most similar. Also, the
results indicated closeness of cortical activities between non-painful
and painful laser stimulations as well as painful laser and thermal
stimulations (see Fig. 69). These results suggest that laser might induce
similar beta responses whether it evoked painful or non-painful feelings but non-painful laser might share presumably non-nociceptive
somatosensory information with mechanical stimulation whereas
painful laser shared presumably nociceptive somatosensory information with thermal stimulation. We expect that further information
theoretical analyses may reveal more details about somatosensory
information encoded in cortical rhythms induced by laser.
References
1. Jun J-H, Park J-R, Kim S-P, Bae YM, Park J-Y, Kim H-S, Choi S, Jung SJ, Park
SH, Yeom D-I. Laser-induced thermoelastic effects can evoke tactile
sensations. Sci Rep. 2015;5.
2. Pfurtscheller G, Lopes Da Silva FH. Event-related EEG/MEG synchronization and desynchronization: basic principles. Clin Neurophysiol.
1999;110(11):1842–57.
3. Celisse A, Robin S. Nonparametric density estimation by exact leave-pout cross-validation. Comput Stat Data Anal. 2008;52(5):2350–68.
P126
Phase synchronization of alpha activity for EEG‑based personal
authentication
Jae‑Hwan Kang1, Chungho Lee1, Sung‑Phil Kim1
1
Department of Human and Systems Engineering, Ulsan National
Institute of Science and Technology, Ulsan
Correspondence: Sung‑Phil Kim ‑ spkim@unist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P126
There has been a growing interest in the EEG-based biometric system as an alternative approach to personal authentication (PA). In
this study, we focused on the potentialities of functional connectivity,
especially phase synchronization in the alpha rhythm represented by
the phase locking value (PLV) as a novel EEG signature for PA. We analyzed an EEG dataset of 39 trials from 7 subjects who participated in
the 5–7 sessions repeatedly on different days. In the sessions, a total
of 16 EEG signals were acquired by a portable EEG device, when the
subjects were in a resting state with their eyes closed for 2 min. The
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 70 Overall characteristics of alpha phase synchronization for
PA. A The upper triangle of association matrix indicates the PLV
calculated by grand mean phase coherence. The lower triangle of
association matrix indicates the its CI values in pairs. B Topographical
connections with the 12 top-ranked CI from the lower triangle of a
characteristics of alpha phase synchronization were estimated by
the following procedures. (1) The alpha rhythm was extracted from
the EEG signals using a band-pass filter of 8–13 Hz. (2) We randomly
selected 20 2-s time segments of the alpha rhythm and calculated
the mean phase coherence [1] between channels within each time
segment. From all possible pairs of 16 EEG channels, a total of 120
mean alpha phase coherence values were extracted. 3) From these
mean alpha phase coherence values, we calculated a criteria index (CI)
of each of them where the CI calculated the ratio of an inter-subject
variability to an intra-subject variability, which was developed to discriminate critical EEG features for PA in our previous study [2]. 4) Using
the mean alpha phase coherence and its CI values, we constructed
the association matrix of phase coherence and extracted the 12 topranked CI connections (Fig. 70). The topographical result showed that
there were apparently two functional connectivity networks of the
alpha rhythm in the brain for PA. The first network was distributed over
the anterior regions including the pre-frontal, frontal and left central
regions. The second one was located in the posterior region covering
the occipital region. It should be noted that two regions have been
well known as main sources of the alpha rhythm from many studies on
EEG alpha rhythms. Our results suggest an important role of the alpha
rhythm in the EEG-based biometrics system.
Acknowledgements: This work was supported by Institute for Information and communications Technology Promotion (IITP) grant
funded by the Korea government (MSIP) (R0190-15-2054, Development of Personal Identification Technology based on Biomedical Signals to Avoid Identity Theft).
References
1. Mormann F, Lehnertz K, David P, Elger CE. Mean phase coherence as
a measure for phase synchronization and its application to the EEG of
epilepsy patients. Phys D. 2000;144:358–69.
2. Kang J-H, Lee C, Kim S-P. EEG Feature Selection and the use of Lyapunov
exponents for eeg-based biometrics. In: IEEE international conference on
biomedical and health informatics, Las Vegas, NV, USA; 2016. p. 1–4.
P129
Investigating phase‑lags in sEEG data using spatially distributed
time delays in a large‑scale brain network model
Andreas Spiegler1, Spase Petkoski1,2, Matias J. Palva3, Viktor K. Jirsa1
1
INSERM UMR 1106 Institut de Neurosciences des Systèmes ‑
Aix‑Marseille Université, Marseille, France; 2Aix‑Marseille Université, CNRS,
ISM UMR 7287, 13288, Marseille, France; 3Neuroscience Center, University
of Helsinki, Helsinki 00014, Finland
Correspondence: Andreas Spiegler ‑ Andreas.Spiegler@univ‑amu.fr
BMC Neuroscience 2016, 17(Suppl 1):P129
On a large scale, the brain appears as a network composed of white
matter tracts connecting brain areas within and between cerebral
hemispheres. The finite transmission speed delays the interaction of
Page 77 of 112
Fig. 71 A Model with local (left) and long-range connections (right).
B, C Averaged tract lengths and weights from 4 connectomes B Joint
distribution, and C histogram of weighted lengths for intra- and interhemispheric links. D Sketch of the spatial delay structure. E Phase-lag
distributions (top) and phase-lags between areas (bottom rows)
areas via these pathways. The delays are on the same scale as the brain
oscillates, that is, 10–250 ms [1], and have been suggested to play a
role in the functional organization. One potential key mechanism is
synchronization [2], which could explain the phase-lags of brain signals. In humans, stereotactical EEG (sEEG) revealed frequency-specific
inter-areal synchronization often associated with nearly zero phaselag, or with variable phase-lags between ±π.
With the advance of non-invasive imaging techniques, large-scale
modeling of the entire brain has become feasible using realistic connectivity and time delays [3], Fig. 71A. From structural and diffusion
MRI, we obtained human connectomes composed of the strength and
length of connections among 68 cortical areas. We approximated the
bimodal tract length distribution (Fig. 71B) by Dirac deltas (Fig. 71C).
Intra-hemispheric connections are in the 1st, and inter-hemispheric
ones are in the 2nd mode. The delay of a connection was determined
from its length divided by the speed of 5 m/s. The Kuramoto phase
oscillator described the activity in each area. The phase difference of
areas was analyzed and compared with the map of inter-areal phase
lags obtained from resting state sEEG of epileptic patients.
The model of fixed oscillators (e.g., f = 20 Hz) switched from global
incoherence to alternating in- and anti-phase coherence with increasing coupling strength. Increasing the natural frequencies for constant
coupling resulted in alternating switching from in- to anti-phase
coherence, but also to incoherence. Intra-hemispheric links were inphase (phase-lags ~0), and inter-hemispheric links were either in- or
anti-phase (±π), see clusters in Fig. 71E. Links among areas of low instrength (sum of all the weights for that node) showed flatly distributed phase-lags. For f = 20 Hz, we found the phase-lags in the sEEG
in the regime of in- and anti-phase coherence in the model, Fig. 71E.
We demonstrated that it is not simply the connectivity strength that
matters in oscillatory large-scale brain networks, but time delays are of
equal importance. The spatial structure in the time delays is reflected
in the clustering of phase-lags. The model captured the statistics of the
phase-lags as observed in the experimental data. The phase-lag structure of links at f = 20 Hz is explained in the model by a spatial organization of in- and anti-phase coherence.
References
1. Buzsáki G, Draguhn A. Neuronal oscillations in cortical networks. Science.
2004;304(5679):1926–29.
2. Varela F, Lachaux J, Rodriguez E, Martinerie J. The brainweb: phase
synchronization and large-scale integration. Nat Rev Neurosci. 2001;2(4):
229–39.
3. Deco G, Jirsa V, McIntosh AR, Sporns O, Kötter R. Key role of coupling,
delay, and noise in resting brain fluctuations. Proc Natl Acad Sci USA.
2009;106(25):10302–07.
P130
Epileptic seizures in the unfolding of a codimension‑3 singularity
Maria L. Saggio1, Silvan F. Siep1, Andreas Spiegler1, William C. Stacey2,
Christophe Bernard1, Viktor K. Jirsa1
1
INSERM UMR 1106 Institut de Neurosciences des Systèmes ‑
Aix‑Marseille Université, Marseille, France; 2Department of Neurology,
BMC Neurosci 2016, 17(Suppl 1):54
2
Department of Biomedical Engineering, University of Michigan, Ann
Arbor, MI 48109, USA
Correspondence: Maria L. Saggio ‑ marisa.saggio@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P130
Seizures can arise under a variety of conditions. Despite this fact, there
are invariant features resulting in a characteristic electrophysiological
signature. Investigations of these universal properties lead to a classification of a planar description of point-cycle fast-slow bursters [1] and
a taxonomy of seizures [2]. The phenomenological model, the epileptor [2], is able to reproduce the main features of the predominant class
of human seizure (~80 % of all cases), according to data from epileptic
patients. We aim at generalizing this model to include other bursting
classes of the taxonomy.
We extended the work by [3] on bursters, living in the unfolding of
high codimension singularities, and systematically investigated the
unfolding of the codimension-3 degenerate Takens–Bogdanov bifurcation (focus, elliptic, saddle and cusp cases) [4]. The biological relevance of this codimension-3 bifurcation has been highlighted by other
authors, and several classes appearing in the context of neuronal
spiking have been identified in its unfolding (e.g., [5]). However, a systematic search in this unfolding for all the planar point-cycle bursting
classes predicted by [1] was still missing. The existence of 16 bursters
of the type slow-wave (self-oscillating slow subsystem), and 16 of type
hysteresis-loop (slow-subsystem oscillating thanks to feedback from
the fast one) was predicted.
We could find all slow-wave bursters in the unfolding together with
seven of the hysteresis-loop ones. With regard to these hysteresis-loop
bursters, we propose a model able to reproduce each of them depending only on the initial and final points of the path in the unfolding’s
parameter space. This model is based on the known normal form of the
codimension-three bifurcation [4], therefore we can readily describe
the role of all its variables and how the tuning of its parameters affects
the models activity. We found that the codimension-three model incorporates not only the repertoire (80 % of seizure) of the model proposed
by [2] but also the classes that account for the remaining 20 % of seizures. Moreover, based on an ultra-slow modulation of the bursting
path (see also [6]) in the model, possible transitions between bursting
classes and, more importantly, transitions to regimes (in the parameter
space) where bursting behavior is not possible at all could be predicted.
These predictions could be tested using data from epileptic patients for
whom different types of seizures coexist.
Overall, the main points of the present work are threefold: (i) a model
description comprising the complete set of slow-wave bursters and
seven (out of 16) hysteresis-loop bursters predicted by Izhikevich
[1], (ii) a generalization of the model proposed by [2] to include the
missing seizure types found in human data and to make prediction
about their robustness, (iii) a framework to investigate the coexistence of different seizure types in the same patient and the transitions between them. The possibility of describing different seizure
types with a unique model, thus with a unique set of variables and
parameters, will facilitate the search for physiological correlates and
treatments.
References
1. Izhikevich EM. Neural excitability, spiking and bursting. IJBC.
2000;10(6):1171–266.
2. Jirsa VK, Stacey WC, Quilichini PP, Ivanov AI, Bernard C. On the nature of
seizure dynamics. Brain. 2014;137(8):2210–30.
3. Golubitsky M, Josic K, Kaper TJ. An unfolding theory approach to bursting
in fast-slow systems. In: Krauskopf B, Broer HW, Vegter G, editors. Global
analysis of dynamical systems; 2001. p. 227–308.
4. Dumortier F,Roussarie R, Sotomayor J, Żaładek H. Bifurcations of planar
vector fields—nilpotent singularities and abelian integrals. Berlin:
Springer; 1991.
5. Osinga HM, Sherman A, Tsaneva-Atanasova K. Cross-currents between
biology and mathematics: the codi-mension of pseudo-plateau bursting.
DCDS-A. 2012;32(8):2853–77.
6. Franci A, Drion G, Sepulchre R. Modeling the modulation of neuronal bursting: a singularity theory approach. SIAM J Appl Dyn Syst.
2014;13(2):798–829.
Page 78 of 112
P131
Incremental dimensional exploratory reasoning
under multi‑dimensional environment
Oh‑hyeon Choung1, Yong Jeong1
1
Department of Bio and Brain Engineering, KAIST, Daejeon, 34141, South
Korea
Correspondence: Oh‑hyeon Choung ‑ iohyeonki@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P131
In our daily life, we encounter thousands of complex problems, which
are not ‘one dimensional’. However, multi-dimensional problems were
known to suffer from “curse of dimensionality” [1]. Therefore, the
researches of reward learning and goal-directed behavior were mostly
focused on single dimensional environment for a decade [2]. Even a
few researches on multi-dimensional tasks was emphasizing that
human representation learning is done by reducing the dimensionality, but not focusing on multiple compositional reasoning under multidimensional environment [3].
Here, the multi-dimensional decision task was conducted (Fig. 72A,
B) and the framework of Reinforcement Learning (RL) was used for
analysis. We investigated that the reasoning under multi-dimensional
environment is processed in incremental order, rather than one-shot
learning. Also, the exploration of the best strategy occurs depends
more on internal value, that is exploring under low value and exploiting under high value (softmax decision rule) rather than random
exploration (randomized ɛ-greedy algorithm). Functional MRI were
taken on each subject, while conducting the behavioral task. Brain
regions of the incremental learning and the value sensitive explorative
behavior will be verified.
We demonstrated that incremental learning rule can explain the
multidimensional reasoning process better than other models
(Fig. 72C–E). This result indicate that people deal with the complicated
multi-dimensional problem, we solve them by adding dimensional
information one by one.
References
1. Sutton RS, Barto AG. Reinforcement learning: An introduction. MIT Press;
1998.
2. Niv Y, Daniel R, Geana A, Gershman SJ, Leong YC, Radulescu A, Wilson
RC. Reinforcement learning in multidimensional environments relies on
attention mechanisms. J Neurosci. 2015;8145–57.
3. Lee SW, Shimojo S, O’Doherty JP. Neural computations underlying
arbitration between model-based and model-free learning. Neuron.
2014;687–99.
Fig. 72 Multidimensional decision making task design and model
comparison. A Multidimensional decision making task schematics.
B Systemic structure of the task. C The result of model comparison,
proposed model has significantly high accuracy on prediction. D, E
The models’ prediction accuracy of proposed model (D) and naïve
model (E)
BMC Neurosci 2016, 17(Suppl 1):54
Page 79 of 112
P132
A low‑cost model of eye movements and memory in personal
visual cognition
Yong‑il Lee1,2, Jaeseung Jeong1,2
1
Department of Bio and Brain Engineering, College of Engineering, Korea
Advanced Institute of Science and Technology (KAIST), Daejeon, 34141,
South Korea; 2Program of Brain and Cognitive Engineering, College
of Engineering, Korea Advanced Institute of Science and Technology
(KAIST), Daejeon, 34141, South Korea
Correspondence: Jaeseung Jeong ‑ jsjeong@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P132
P133
Complex network analysis of structural connectome of autism
spectrum disorder patients
Su Hyun Kim1,2, Mir Jeong1, Jaeseung Jeong1,2
1
Department of Bio and Brain Engineering, College of Engineering,
Korea Advanced Institute of Science and Technology (KAIST), Daejeon,
34141, Korea; 2Program of Brain and Cognitive Engineering, College
of Engineering, Korea Advanced Institute of Science and Technology
(KAIST), Daejeon, 34141, Korea
Correspondence: Jaeseung Jeong ‑ jsjeong@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P133
Eye movements are the most useful and clearest signal of our body to
understand cognitive process and memory mechanism. Because it is
convenient to measure the stream signal and quantify. Other sensory
inputs, like auditory, gustatory, olfactory, and kinesthetic stimuli, are
hard to estimate, but visual stimuli are easy by the eye-tracker [1]. And
people are primarily visually oriented. Every day, people get over 80
percent of information from their own eye.
Until now, eye movements data has been measured and analyzed by
very expensive eye-trackers mostly. The major companies’ devices price,
including SMI, Tobii and EyeLink, are at least 10,000$ with their analysis tool SW. Of course, there have been many substitution trials in open
SW and open HW area [2]. But their performance is certainly lower than
major brands. Also, their data representation doesn’t have standard.
Therefore, open SWs are difficult to apply for other utility services and
scientific researchers have hesitated to use them to analyze result data
to understand complex cognitive process. A few research results, which
is investigating the relation between eye movement and cognitive
model based on open eye-tracker platform, have reported by this time.
However, eye-tracker is not only for science, and other areas need the
usability of eye movement, for instance, UI/UX, healthcare, driving,
game, learning consulting, TV viewer rating, market research and so
on [3]. If there is a more general and low-cost eye-tracker which is confirmed cognitive model, above areas would be effective and we could
do better decision making. This research implements a low-cost eyetracker using a front camera (webcam) and a pin camera (Fig. 73, If the
pc or laptop has laptop has a front camera, it doesn’t need more pin
camera). The implementation includes the auto detection and classifying of useful memory based on eye movement of visual information
on the device’s display. To do this function, the camera measures the
saccade variation spectrum, as the X–Y axis acceleration, and categorize individual pattern while the user is taking train session. It is developed using OpenCV library and C#. XLabs Inc., already has made the
gaze/head tracker using front camera without cognitive pattern analysis [4]. In the future, we will try this function on the mobile devices,
which are cellular phone, tablet pc, and game interface. These devices
have more sensors, like GPS, illumination, and activity accelerator.
Combination of sensors input would make more precise prediction for
memory cognition.
Background Human connectome which is the map of full connection of neuronal network in the human brain exhibits the characteristics of complex network [1]. The human connectome is known to
be wired in a way that neurons efficiently transmit and communicate
information. Statistical measures of complex network describe the
topological features of a certain network and this enables researchers
to compare effectiveness of information processing within a network.
Autism spectrum disorder (ASD) subjects exhibit repetitive behaviors,
impaired social communication skills, and sensory problems. Those
symptoms of neurodevelopmental disorder is doubted to be originated from genetic causes [2]. Also, recent investigations find that ASD
is a ‘connection problem’. But still exact cause of ASD is unknown. The
aim of the research to be conducted is to reveal the genetic cause of
(ASD) by the statistical analysis of network measures of ASD patients
and normal groups’ structural connectome data using diffusion tensor
imaging (DTI).
Methods The method to be used in the research is to compare the
network measure values of various subject groups’ connectome and
relate the difference to the genetic mutations in common. DTI data
describes the neuronal connection of the brain regions in the mesoscale level. Structural connectome that is constructed from DTI
information.
Expected result is that there are genotypic changes of genes which
affect development of neuronal connection in ASD subjects. This finding will shade a new light on the investigation of ASD diagnosis and
treatment.
References
1. Wedel M, Pieters R: Eye tracking for visual marketing. Now Publishers Inc;
2008.
2. Dalmaijer ES, Mathôt S, Van der Stigchel S. PyGaze: an open-source,
cross-platform toolbox for minimal-effort programming of eyetracking
experiments. Behav Res Methods. 2014;46(4):913–21.
3. Pannasch S, Helmert JR, Velichkovsky BM. Eye tracking and usability research: an introduction to the special issue. J Eye Mov Res.
2008;2(4):1–4.
4. https://xlabsgaze.com/about/.
Fig. 73 Low-cost eye movement tracker using front cameras on the
each devices
References
1. Bullmore E, Sporns O. Complex brain networks: graph theoretical analysis
of structural and functional systems. Nat Rev Neurosci. 2009;10(3):186–98.
2. Freitag CM, Staal W, Klauck SM, Duketis E, Waltes R. Genetics of autistic
disorders: review and clinical implications. Eur Child Adolesc Psychiatry.
2010;19(3):169–78.
P134
Cognitive motives and the neural correlates underlying human
social information transmission, gossip
Jeungmin Lee1,2, Jaehyung Kwon1, Jerald D. Kralik1,2, Jaeseung Jeong1,2
1
Department of Bio and Brain Engineering, College of Engineering, Korea
Advanced Institute of Science and Technology (KAIST), Daejeon, 34141,
Republic of Korea; 2Program of Brain Engineering, College of Engineering,
Korea Advanced Institute of Science and Technology (KAIST), Daejeon,
34141, Republic of Korea
Correspondence: Jaeseung Jeong ‑ jsjeong@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P134
Gossip is a specific example of human conversation containing social
factors and has been considered as malicious, useless idle-talk by general population. Several researchers have suggested the role of gossip
as social police that control the members of social groups to behave
cooperative rather than selfish. However, there is not enough data to
explain the actual cognitive motives that drive people to spread gossip. Throughout this study, human gossiping behavior is defined as
transmission of social information about an absent third-party (i.e.
the target of the gossip). In order to define the types of gossip, various scenarios containing social information are divided into 48 different categories by the third-party identity, valence and contents.
Big five personality inventory, prosocial personality battery, cultural
orientation scale and moral foundations questionnaire were used to
BMC Neurosci 2016, 17(Suppl 1):54
measure personal traits that may influence the gossiping behavior of
individuals. We found out that people, regardless to the scores of their
personal traits, tend to spread gossip about in-group and celebrities
more than out-group members. We also found out that positive gossip about in-group members is spread with significantly higher rates
than in-group negative gossip, whereas the spread pattern was the
opposite when the gossip is about celebrities. With such findings, we
conducted fMRI study using in-group and out-group gossip with positive and negative valence. Increased activity in various brain regions
including medial frontal gyrus, dorsolateral prefrontal cortex and precuneus was found when participants made decisions whether or not
to spread gossip. With the obtained data, we tried to construct a computational model that may be used for classification of spread gossip.
P135
EEG hyperscanning detects neural oscillation for the social
interaction during the economic decision‑making
Jaehwan Jahng1,2, Dong‑Uk Hwang3, Jaeseung Jeong1,2
1
Department of Bio and Brain Engineering, College of Engineering, Korea
Advanced Institute of Science and Technology (KAIST), Daejeon, 34141,
South Korea; 2Program of Brain and Cognitive Engineering, College
of Engineering, Korea Advanced Institute of Science and Technology
Page 80 of 112
(KAIST), Daejeon, 34141, South Korea; 3Division of Computational
Mathematics, National Institute for Mathematical Sciences (NIMS),
Daejeon, 34047, South Korea
Correspondence: Jaehwan Jahng ‑ jahngjh.627@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P135
Social interaction is an important feature of the economic exchange.
However, little is known about the varying neural mechanism during
the economic decision-making depending on the different degrees
of the social interaction. In this study, we used an iterated version of
Prisoner’s dilemma game (PDG) with an EEG hyperscanning to investigate how the presence of face-to-face interaction modulates the social
interactions and in turn the aspects of an economic decision-making.
Participants played the game either face-to-face (FF) or face-blocked
(FB). On the behavioral level, face-to-face interaction led both participants to choose cooperative strategies more often. On the neural
level, FF groups showed significantly different alpha power during
the first 0.5 s after seeing each outcome compared with FB groups in
right temporo-parietal region. By computing the phase locking value
(PLV), we measured the brain synchrony and found that the inter-brain
phase synchronies across right temporo-parietal area were significantly associated with both the group differences and strategical differences of both players (Fig. 74). These results suggest that inter-brain
alpha synchronies across right temporo-parietal area might serve
as an implicit neural marker for both the social interaction level and
intention to either cooperate or defect. Moreover, our results warrant
the future hyperscanning studies on the social interactions of autism
spectrum disorder (ASD) patients as all neural substrates revealed are
known to be deeply associated with their social traits.
Acknowledgements: This research was supported by the CHUNG
Moon Soul Research Center for Bio Information and Bio Electronics
(CMSC) in KAIST and a Korea Science and Engineering Foundation
(KOSEF) grant funded by the Korean government (No. 2006-2005399).
The funders had no role in study design, data collection and analysis,
the decision to publish, or the preparation of the manuscript.
P136
Detecting purchase decision based on hyperfrontality of the EEG
Jae‑Hyung Kwon1,2, Sang‑Min Park1,2, Jaeseung Jeong1,2
1
Department of Bio and Brain Engineering, Korea Advanced Institute
of Science and Technology (KAIST), Daejeon, 34141, South Korea;
2
Program of Brain and Cognitive Engineering, Korea Advanced Institute
of Science and Technology (KAIST), Daejeon, 34141, South Korea
Correspondence: Jae‑Hyung Kwon ‑ jh2393@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P136
Fig. 74 A Brain synchrony analyses. Intra-brain and inter-brain phase
synchronies in alpha band [0.5, 1] s. Links between electrodes means
that the phase activities there are synchronized. All synchronies
here were higher in FF groups than in FB groups (gray line). Blue line
denotes the synchronies that were higher in CC epochs compared
with DD epochs of FF groups (CC > DD) whereas red line denotes
the synchronies that were higher in DD epochs compared with CC
epochs of FF groups (CC < DD). Intra-brain synchronies are drawn
in both brains and only one of each pair of inter-brain synchronies
are drawn. Significant level was at p < 0.05, Bonferroni corrected. B
Magnitudes of phase synchronies that showed significant strategical
differences. These correspond to the links depicted as blue and red
line in A. * p < 0.05; ** p < 0.01, Bonferroni corrected
Understanding and predicting purchase decision process is one of the
fundamental issues in economics, marketing, decision sciences, yet
an easily accessible means to monitoring purchase decisions has not
been developed yet [1–3]. Using event-related functional fMRI, such
purchase behavior with a shopping task was investigated [4] but it
has limit on potential practical uses due to the cost and portability of
the MRI. The electroencephalogram (EEG) has been suggested to have
many advantages for applications in marketing due to its relatively low
cost, portability, and high temporal resolution. The aim of the current
study was to determine the possibility of the EEG as a tool for detecting and predicting purchase decision in potential consumers. Twentythree participants were recruited to record their EEGs as they saw the
pictures of products followed by the products’ prices and made the
choice of whether to buy them or not. We estimated the power spectra and approximate entropy (ApEn), an information-theoretic measure to quantify the complexity [5], of their EEGs and compared them
for purchase and non-purchase trials. The support vector machine
(SVM) method was to predict their purchase decisions. We found that
the relative spectral powers and ApEn values of the EEG significantly
differed between purchase and non-purchase trials, in particular
frontal regions. SVM could distinguish and predict purchase and nonpurchase decisions based on the spectral powers and ApEn values of
the EEGs in frontal regions prior to the decision moment with a high
accuracy (>87 %). This finding suggests that relatively inexpensive,
BMC Neurosci 2016, 17(Suppl 1):54
portable EEG recording technique has great potential as a neural predictor of purchase behavior in neuromarketing and neuroeconomics.
References
1. Lee N, Broderick AJ, Chamberlain L. What is “neuromarketing”? A
discussion and agenda for future research. Int J Psychophysiol.
2007;63:199–204.
2. Mirja Hubert PK. A current overview of consumer neuroscience. J Consum Behav. 2008;7:272–92.
3. Ariely D, Berns GS. Neuromarketing: the hope and hype of neuroimaging
in business. Nat Rev Neurosci. 2010;11:284–92.
4. Knutson B, Rick S, Wimmer GE, Prelec D, Loewenstein G. Neural predictors
of purchases. Neuron. 2007;53:147–56.
5. Gu F, Meng XIN, Shen E, Cai Z. Can we measure consciousness with EEG
complexities? Int J Bifurc Chaos. 2003;13:733–42.
P137
Vulnerability‑based critical neurons, synapses, and pathways
in the Caenorhabditis elegans connectome
Seongkyun Kim1, Hyoungkyu Kim1, Jerald D. Kralik1, Jaeseung Jeong1
Department of Bio and Brain Engineering, Program of Brain and Cognitive
Engineering, College of Engineering, Korea Advanced Institute of Science
and Technology (KAIST), Daejeon, 34141, South Korea
Correspondence: Jerald D. Kralik ‑ jerald.kralik@raphe.kaist.ac.kr, Jerald D.
Kralik ‑ jsjeong@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P137
Determining the fundamental architectural design of complex
nervous systems will lead to significant medical and technological
advances. Yet it remains unclear how nervous systems evolved highly
efficient networks with near optimal sharing of pathways that yet
produce multiple distinct behaviors to reach the organism’s goals.
To determine this, we investigated the vulnerability of the nematode
Fig. 75 Three critical pathways emerged from the results. A For VB
they were: (1) AVA-based; (2) PVP-based; and (3) RMD → OLL. B Two
of these pathways were again implicated for VE: the AVA-based and
the PVP-based pathways
Page 81 of 112
roundworm Caenorhabditis elegans connectome [1] by attacking each
of 279 individual neurons and 6393 chemical synapses and 890 electrical junctions in the connectome, and quantifying the lethality of the
network in terms of global information processing using graph-theoretic measures: i.e., examining vulnerability with respect to clustering
(C), efficiency (E), and betweenness (B).
The vulnerability analyses, VC, VE, VB, identified 12 critical neurons
and 29 critical synapses that are the most important components for
establishing fundamental network properties. These critical elements
were found to be control elements—i.e., those with the most influence
over multiple underlying pathways. In addition, we found that the critical synapses formed into circuit-level control units, suggesting fractallike control in the connectome. More specifically, three main critical
pathways emerged from the results (Fig. 75A, B).
Conclusions The critical pathways that emerged from our computational analysis provide evidence for (a) the importance of backward
locomotor control, avoidance behavior, and social feeding to the organism; (b) the potential roles of specific neurons whose functions have
been unclear; and (c) both parallel and serial design elements in the
connectome—i.e., specific evidence for a mixed architectural design.
This design structure may be fundamental to nervous systems, providing necessary building blocks for the evolution of higher intelligence.
Acknowledgements:: Supported by the National Research Foundation of Korea (NRF-2013R1A1A2011570).
Reference
1. Altun Z, Hall D. Worm Atlas; 2002. http://www.wormatlas.org.
P138
Motif analysis reveals functionally asymmetrical neurons in C.
elegans
Pyeong Soo Kim1, Seongkyun Kim1, Hyoungkyu Kim1, Jaeseung Jeong1
1
Department of Bio and Brain Engineering, Korea Advanced Institute
of Science and Technology (KAIST), Daejeon 305‑701, South Korea
Correspondence: Jaeseung Jeong ‑ jsjeong@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P138
Majority of animal species have bilaterally symmetrical nervous system. Symmetric and asymmetric features among their morphological
symmetric nervous system have been interesting issue for long time.
The simplest bilaterally symmetrical organism is nematode called
Caenorhabditis elegans. Previously, symmetry for C. elegans has only
been thoroughly studied in morphological and functional manner [1].
According to previous observation, there are 92 bilaterally symmetrical neuronal pairs and remaining 95 neurons are mostly located on
the axis of symmetry. Functionally there are only 2 neuronal pairs that
show asymmetrical gene expression among 92 pairs of symmetrical
neurons. We examined the symmetry of C. elegans nervous network
which has not been studied.
Total of 279 neurons and 2990 links in C. elegans were used. Neurons
were classified into bilaterally symmetrical neurons, unlateral neurons,
and unilateral neurons. According to the neuronal positions, we could
define the symmetry of each individual link and expand that definition to define the symmetry of motif [2]. After defining symmetry of
nervous network, we suggest a novel approach to classify asymmetric
neurons of C. elegans nervous system by examining asymmetric network topology for every node. We defined 5 explicit locally topological parameters for a neuron; (1) the degree is defined as the number
of asymmetric links attached to the neuron, (2) the motif is defined
as distribution of the numbers of asymmetric motifs for a neuron, (3)
the degree ratio is defined as ratio of asymmetric links over totally
attached links to the neuron including both of symmetric links and
asymmetric links, (4) the motif ratio is distribution of the rates for
asymmetric motifs over total motifs including both of symmetric and
asymmetric motifs, and (5) the relative distance is defined by the difference of asymmetric motif fingerprint of bilaterally symmetrical
neurons. Thresholds were defined using mean and standard deviation
(SD) values of asymmetries to find statistically asymmetric components. Neurons with asymmetry value over the threshold were considered as asymmetric neurons (asymmetric neurons > SD from the
BMC Neurosci 2016, 17(Suppl 1):54
mean values). We checked our asymmetric neurons with ASE and AWC
neurons that are only known to show bilaterally asymmetrical gene
expression. As a result, our study suggested that (4) ratio of asymmetric motif and (5) relative distance measures successfully classified ASE
and AWC as asymmetric neurons. Except for ASE and AWC neurons,
BDU, PLM, and PVW neurons are classified asymmetric in both measures. These results could be interpreted that BDU neurons, PLM neurons, and ALN neurons might possess asymmetric features that have
not been discovered.
References
1. Oliver H, Johnston RJ, Chang S. Left–right asymmetry in the nervous system: the Caenorhabditis elegans model. Nat Rev Neurosci.
2002;3(8):629–40.
2. Sporns O, Kötter R. Motifs in brain networks. PLoS Biol. 2004;2(11):e369.
P139
Computational approach to preference‑based serial decision
dynamics: do temporal discounting and working memory affect
it?
Sangsup Yoon1,2, Jaehyung Kwon1,2, Sewoong Lim1,2, Jaeseung Jeong1,2
1
Department of Bio and Brain Engineering, College of Engineering, Korea
Advanced Institute of Science and Technology (KAIST), Daejeon, 34141,
South Korea; 2Program of Brain and Cognitive Engineering, College
of Engineering, Korea Advanced Institute of Science and Technology
(KAIST), Daejeon, 34141, South Korea
Correspondence: Jaeseung Jeong ‑ jsjeong@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P139
When we face the multiple options (such as different taste of chocolates in a box) and choose one by one sequentially, there is no reason
to prefer particular order of choice than any other possible choice
strategies since all the items will eventually be consumed by ourselves.
Recent studies have, however, revealed that there are distinct patterns of choice strategy in this preference-based serial decision tasks.
Interestingly, there are two opposite choice patterns (favorite-first and
favorite-last) in human subjects [1], while non-human animals (rhesus
monkeys) only chose their favorite options at first [2]. Although several
hypotheses for underlying neural mechanisms have been suggested
to explain about how these distinct choice strategies appeared, they
are not directly tested yet.
The goal of the current study was to examine whether temporal
discounting and working memory affect choice strategy of serial
decision-making and if so, to examine how they influence it. To measure the choice strategy, we used the modified version of ‘the sushi
problem’ task [1], which use the pictures of opposite sex as a reward
instead of the sushi [3]. We also measured the temporal discounting
parameters and working memory performance by using the same set
of opposite-sex pictures. The whole pictures were rated by each subject twice before the main experiment, the average rating score were
subsequently used to divide the whole picture set into four groups
based on the difference of subject specific preference. In ‘sushi’ task,
subjects were asked to choose among four options (squares contain
different number of stars; 1 ~ 4), each of which represents following
short presentation (1.5 s) of pictures right after their choice. Each trial
ends when subject choose all of four options, so they couldn’t skip or
miss any options. The temporal discounting parameter was measured
by the subject choice between sooner-small reward and later-larger
reward, in this case, the magnitude of reward was the number of stars
which indicate the subject-specific attractiveness of each pictures, the
delay of reward was relatively shorter (1–30 s) than typical temporal
discounting task since our task offered actual outcome (see the picture) of each choices [4]. The picture version of n-back task was used to
measure the working memory performance.
Consistent with previous studies, we observed distinct patterns of
choice order in ‘sushi’ task, favorite-last strategy was most dominant
(58 %) and favorite-first was second (31 %). We also found the relationship of both temporal discounting and working memory with choice
strategies. The favorite-last group showed significantly lower rate of
temporal discounting and higher performance of working memory
Page 82 of 112
than favorite-first group. The effects of working memory and temporal
discounting parameters on choice strategy were examined by logistic
regression analysis, which revealed how the propensity to discount
future events and the memory effect about recent events predicted
the pattern of serial choice. We also constructed simple computational models using support vector machine and naïve Bayes classifier
to predict their decision patterns based on working memory performance and temporal discounting parameters. We showed that these
computational models successfully predicted preference-based decision patterns.
References
1. Jaeseung J, Younmin O, Miriam C, Jerald DK. Preference-based serial decision dynamics: your first sushi reveals your eating order at the sushi table.
Plos One. 2014;9(5):e96653
2. Kanghoon J, Jerald DK. Get it while it’s hot: a peak-first bias in self-generated choice order in rhesus macaques. Plos One. 2013;8(12):e83814.
3. Itzhak A, Nancy E, Dan A, Christopher FC, Ethan O, Hans CB. Beautiful
faces have variable reward value: fMRI and behavioral evidence. Neuron.
2001;32:537–51.
4. Benjamin YH, Purak CP, Robert OD, Michael LP. Economic principles motivating social attention in humans. Proc R Soc B. 2007;274:1751–56.
P141
Social stress induced neural network reconfiguration affects
decision making and learning in zebrafish
Choongseok Park1, Thomas Miller2, Katie Clements2, Sungwoo Ahn3, Eoon
Hye Ji4, Fadi A. Issa2
1
Department of Mathematics, North Carolina A&T State University,
Greensboro, NC, 27411, USA; 2Department of Biology, East Carolina
University, Greenville, NC, 27858, USA; 3Department of Mathematics, East
Carolina University, Greenville, NC, 27858, USA; 4David Geffen School
of Medicine, UCLA, Los Angeles, CA, 90095, USA
Correspondence: Choongseok Park ‑ issaf14@ecu.edu,
Fadi A. Issa ‑ cpark@ncat.edu
BMC Neuroscience 2016, 17(Suppl 1):P141
In many social species, behavioral mechanisms of how social hierarchies formed and maintained have been studied extensively [1]. However, the neural bases underlying behavioral decisions and dynamics
of neural circuits that permit animals to adapt to changes in social rank
are poorly understood. In this study we focused on two social stress
induced behaviors in zebrafish [the Mauthner cell (M-cell) mediated
startle escape response and swimming behavior] to investigate how
social regulation affects intrinsic cellular and network properties that
result in the behavioral differences between dominant (DOMs) and
subordinate (SUBs) animals. We utilized a non-invasive technique that
allowed us to monitor the activation pattern of the two neural circuits
in freely behaving animals.
High behavioral responsiveness and a low stimulus threshold for
the initiation of escape in M-cell were observed in SUBs while DOMs
showed the quicker habituation to repeated auditory stimulation
compared to SUBs. We also observed that on average SUBs generated significantly less number of swim bursts compared to DOMs.
These results suggest that social status induced stress can modify
the startle plasticity as well as the local swimming circuit. The change
in M-cell’s excitability due to the change in the presynaptic inhibitory
drive may be responsible for the lowered threshold. On the other
hand, the local neural circuits and their intrinsic modulatory components (motor neurons and interneurons) may be configured differently according to social status to produce status-dependent swim
patterns [2].
To test these ideas, we developed a biologically-based mathematical
model whose network architecture is based on recent experimental
data [3]. The model is able to reproduce several hallmarks of social status induced behavioral differences that were experimentally observed
between DOMs and SUBs, as well as some inherent activity patterns.
Changing some intrinsic synaptic and network parameters was sufficient to obtain the transition between DOMs and SUBs activity patterns while maintaining the network architecture.
BMC Neurosci 2016, 17(Suppl 1):54
Recent experiments show that the startle plasticity in M-cell can
be modulated by endocannabinoids, 2-AG [3]. We chose the availability of 2-AG in M-cell as one of main parameters for the simulation, whose dynamics is governed by the intracellular calcium level in
M-cell. Model simulation shows that high behavioral responsiveness
in SUBs results from the increased excitability in M-cell, which can be
interpreted as the reduced inhibitory input to M-cell. To reproduce
less swimming activity in SUBs, the hallmark of social status induced
behavioral difference observed in our experiments, we chose another
intrinsic parameter, the availability of 2-AG in inhibitory interneurons
to represent 2-AG modulated local network property. Model simulation shows that less swimming activity in SUBs is produced by the
increased inhibitory input to the swimming neural circuit via the 2-AG
driven elevated interneuron activity.
Acknowledgements: This work was partially supported by Simons
Foundation Collaboration Grants for Mathematicians (#317566) to CP
and East Carolina University, Department of Biology fund to FAI.
References
1. Bergman TJ, Beehner JC, Cheney DL, Seyfarth RM. Hierarchical classification by rank and kinship in baboons. Science. 2003;302(5648):1234–36.
2. Issa FA, Drummond J, Cattaert D, Edwards DH. Neural circuit reconfiguration by social status. J Neurosci. 2012;32(16):5638–45.
3. Song J, Ampatzis K, Ausborn J, Manira AE. A hardwired circuit supplemented with endocannabinoids encodes behavioral choice in zebrafish.
Curr Biol. 2015;25:2610–20.
P142
Descriptive, generative, and hybrid approaches for neural
connectivity inference from neural activity data
JeongHun Baek1, Shigeyuki Oba1, Junichiro Yoshimoto2,3, Kenji Doya2,
Shin Ishii1
1
Graduate School of Informatics, Kyoto University, Yoshidahonmachi
36‑1, Sakyo, Kyoto, Japan; 2Neural Computation Unit, Okinawa
Institute of Science and Technology Graduate University, 1919‑1
Tancha, Onna‑son, Kunigami‑gun, Okinawa, Japan; 3Graduate School
of Information Science, Nara Institute of Science and Technology,8916‑5
Takayama, Ikoma, Nara, Japan
Correspondence: JeongHun Baek ‑ ku21fang@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P142
Identification of the connectivity between neurons is important not
only for elucidating neural bases for various functions but also for
reconstructing the dynamics emerged in the connectivity. This thesis
considers efficient methods for estimating synaptic connections from
neural activity data. There are two kinds of approaches to neural connectivity inference: analytic one based on descriptive statistics and
reproductive one based on statistical generative models. Analyses
based on descriptive statistics, such as Pearson correlation, can identify neural connectivity based on activity data, with low computational
cost. It, however, cannot reproduce the dynamic behaviors of the
underlying connectivity, and hence, it is not suited for simulating the
identified network.
Reproductive approach based on statistical generative models, such
as generalized linear model, can naturally simulate the dynamic
behaviors of the identified network, once we determine the network
parameters from activity data. Contrary to this advantage, the computational cost of reproductive approach is often much heavier than analytic one. To utilize the preferable characters of the two approaches, in
this study, we propose a hybrid approach of using a descriptive statistic for prescreening of existing connections and then performing generative model inference for dynamic model construction. We applied
the hybrid approach to artificially generated spike data of various network sizes.
Results and conclusions Figure 76A shows the accuracy of functional
connectivity analysis, in terms of ROC-AUC of binary classification,
presence/absence of connectivity, where we see the hybrid approach
performed slightly worse than the GFAM10. Note that the hybrid
approach performed almost same with GFAM10 when the number
Page 83 of 112
Fig. 76 A Comparison of the prescreening accuracy in terms of
ROC-AUC value (higher is better accuracy). B Comparison of the computation time. ‘GFAM10 [1]’, regarded as the original method, denotes
generative functional additive model which is extended version of
generalized linear model. ‘Correlation-GFAM10’ denotes a hybrid
approach which performs the Pearson correlation for prescreening
and then performs GFAM10
of neurons was 2000. Figure 76B shows the computation time of the
two methods, where we see that GFAM10 took five times as much time
as the hybrid approach. Our hybrid approach successfully reduced
computational time, into about one-fifth of that of the sole reproductive approach based on the GFAM, while maintaining the estimation
accuracy of the response functions within the identified functional
connectivity.
Reference
1. Song D, Wang H, Tu CY, Marmarelis VZ, Hampson RE, Deadwyler SA,
Berger TW. Identification of sparse neural functional connectivity using
penalized likelihood estimation and basis functions. J Comput Neurosci.
2013;35(3):335–57.
P145
Divergent‑convergent synaptic connectivities accelerate coding
in multilayered sensory systems
Thiago S. Mosqueiro1, Martin F. Strube‑Bloss2, Brian Smith3, Ramon
Huerta1
1
University of California San Diego, La Jolla CA, USA; 2Biocenter University
of Würzburg, Würzburg, Germany; 3School of Life Sciences, Arizona State
University, Tempe, AZ, USA
Correspondence: Brian Smith ‑ brian.h.smith@asu.edu
BMC Neuroscience 2016, 17(Suppl 1):P145
A central dogma in perception postulates that a minimal number of
higher-order neurons provide the coding basis required for decision
making and survival [1]. However, sensory information must travel
through several neural layers before converging onto a smaller number of neurons in a premotor decision layer [2]. This multi-layered
processing and convergence induces a time lag between peripheral
input and adaptive behavior, which is inconsistent with the need for
reaction speed. We propose that the divergent–convergent organization often occurring in multilayered neuropils enhances processing
speed. Insect olfactory processing is a good model for investigating
perceptual timing [3], where effective classification in the 4th layer
‘anticipates’ classification in input layers by 50 ms (Fig. 77A, B) [4].
Here we show that this anticipation emerges from divergent-convergent connectivity and the relative sizes of the layers, which rapidly amplifies subtle input signals and improves precision (Fig. 77C).
We reproduced experimental results of peak classification in MBONs
anticipating PNs by 50 ms on average (Fig. 77D). This becomes more
pronounced as the KC layer grows, although increased noise is also
observed. For an oversized KC layer, thus, this anticipation becomes
lower and the signal is eventually destroyed by the emphasized
noise. Interestingly, the key feature to this anticipation is indeed the
ratio between KCs and PNs, showing that larger brains may balance
these populations to achieve jointly higher pattern recognition capabilities and fasts discrimination times. We have analyzed fast coding
BMC Neurosci 2016, 17(Suppl 1):54
Page 84 of 112
2
Department of Nonlinear Dynamics and Complex Systems, Institute
of Computer Science, The Czech Academy of Sciences, Prague, 182 07,
Czech Republic; 3National Institute of Mental Health, Klecany, 250 67,
Czech Republic
Correspondence: Michal Hadrava ‑ hadrava@cs.cas.cz
BMC Neuroscience 2016, 17(Suppl 1):P146
Fig. 77 Early discrimination of stimulus in the MBs. A Recordings
of PNs and MBONs activities from untrained honey bees to odor
stimulation. At t = 0 s (green bar), an odor stimulation is presented. B
Connectionist blueprint of the MBs, emphasizing synapses and population size. Note the divergence present between PN and KC layers,
followed by convergence onto MBONs. C We reproduced in silico the
early response (blue bar) of MBONs in the vertical lobe with respect
to the PNs (orange bar) using spiking neuron networks. D Time differences in our simulations for each experiment repetition. Difference in
response time between is on average 50 ms (one tail Mann–Whitney
test, p < 0.025)
properties of fan-out/fan-in structures that are ubiquitous in the
brain. We developed a model to reproduce experimental data and
analyze the optimal reaction times of the network model, finding a
balance between fast information transmission and high accuracy in
pattern recognition. Our contribution improves understanding of the
role of divergent convergent feedforward networks on the stability of
fast and accurate decision-making.
References
1. Barlow HB. Single units and sensation: a neuron doctrine for perceptual
psychology? Perception. 2009;38:371–94.
2. Shepherd GM. The synaptic organization of the brain. Oxford: Oxford
Press; 2003.
3. Mosqueiro TS, Huerta R. Computational models to understand decision
making and pattern recognition in the insect brain. Curr Opin Insect Sci.
2014;6:80–5.
4. Strube-Bloss MF, Herrera-Valdez MA, Smith BH. Ensemble response in
mushroom body output neurons of the honey bee outpaces spatiotemporal odor processing two synapses earlier in the antennal lobe. PLoS
One. 2012;7:e50322.
P146
Swinging networks
Michal Hadrava1,2,3, Jaroslav Hlinka2,3
1
Department of Cybernetics, Faculty of Electrical Engineering, Czech
Technical University in Prague, Prague, 166 27, Czech Republic;
Nature is a powerful illusionist who, unfortunately for life sciences,
hates revealing her secrets. One of her most rewarding tricks involves
interconnecting a bunch of non-oscillatory neurons in such a way that
they collectively behave like an oscillator [1]. Contemporary neuroscience strives to decipher this magic and does so not out of mere curiosity: the trick can go wrong, causing myriads of neurons to march
to the deadly rhythm of epileptic seizure. It is of vital importance to
determine which connectivity patterns promote and suppress epileptiform activity if surgery is to be effective when nothing else is [2]. On
a lighter note, oscillatory dynamics explains many aspects of musical
experience [3–5]. A question comes up again and again in ethnomusicological discourse as to whether these aspects are learned or not.
In the context of oscillatory dynamics on networks, we might ask
whether a single connectivity leads to the emergence of dynamics
relevant to each musical culture or a different (learned) connectivity is
at work each time. In conclusion, establishing a link between connectivity and oscillatory dynamics on networks seems to be an important
problem with repercussions in such diverse fields as epileptology and
ethnomusicology.
The mainstream approach to the problem can be characterized as
follows: first, choose a dynamical model of single unit—e.g. neuron,
synapse, or population thereof. Next, connect units of the selected
type(s) in a network. Finally, study the effect of connectivity parameters on the global dynamics analytically, computationally, or using
a combination of both. The major drawback of any analysis performed in this way is that the validity of its results is put in doubt
whenever that of the single unit model is. Needless to say, none of
the ever-growing variety of models has gained a wide acceptance
yet. The mainstream approach could be dubbed the “object-oriented”
one. The alternative approach, advocated by category theorists and
adopted by us, could be called the “relational” one: instead of analysing a particular dynamical system, one investigates a whole class
of dynamical systems on a particular manifold characterized only by
its relations to classes of dynamical systems on different manifolds.
This latter approach is epitomized by a recently introduced algebraic
structure [6] which relates global network dynamics to its connectivity. We are currently trying to prove the existence of global periodic
solutions in selected classes of simple networks with a given structure
using this new theory.
Acknowledgements: This work was supported by the Grant Agency
of the Czech Technical University in Prague, grant No. SGS14/192/
OHK3/3T/13, the Czech Science Foundation Project No. P30314-02634S, and the Czech Health Research Council Project No.
NV15-29835A.
References
1. Buzsáki G. Rhythms of the brain. New York: Oxford University Press; 2006.
2. Dixit AB, Banerjee J, Tripathi M, Chandra PS. Presurgical epileptogenic
network analysis: a way to enhance epilepsy surgery outcome. Neurol
India. 2015;63(5):743–50.
3. Cartwright JHE, Gonzalez DL, Piro O. Pitch perception: a dynamicalsystems perspective. Proc Natl Acad Sci USA. 2001;98(9):4855–59.
4. Large EW: A dynamical systems approach to musical tonality. In: Huys R,
Jirsa VK, editors. Studies in computational intelligence: nonlinear dynamics in human behavior, vol 328. Berlin: Springe; 2011. p. 193–211.
5. Large EW, Snyder JS. Pulse and meter as neural resonance. In: DallaBella S,
Kraus N, Overy K, Pantev C, Snyder JS, Tervaniemi M, Tillmann B, Schlaug
G, editors. Neurosciences and music III: disorders and plasticity, vol 1169;
2009. p. 46–57.
6. Lerman E, Spivak DI. An algebra of open continuous time dynamical
systems and networks. arXiv:1602.01017v1 [math.DS].
BMC Neurosci 2016, 17(Suppl 1):54
P147
Inferring dynamically relevant motifs from oscillatory stimuli:
challenges, pitfalls, and solutions
Hannah Bos1, Moritz Helias1,2
1
Institute of Neuroscience and Medicine (INM‑6) and Institute
for Advanced Simulation (IAS‑6) and JARA BRAIN Institute I, Jülich
Research Centre, 52425 Jülich, Germany; 2Department of Physics, Faculty
1, RWTH Aachen University, 52074 Aachen, Germany
Correspondence: Hannah Bos ‑ h.bos@fz‑juelich.de
BMC Neuroscience 2016, 17(Suppl 1):P147
Applications of oscillatory stimuli in optogenetical studies have been
used to gather evidence that γ oscillations are generated by the
interaction of inter-neurons (also termed the inter-neuron γ or ING
mechanism) [1, 2]. We elaborate the pitfalls of inferring the origin of
the oscillation from absolute (response spectra) as well as relative
(power ratios) changes in spectra of neural activity induced by oscillatory input. We consider minimalistic models that isolate the difficulties and limitations arising in the interpretation of response spectra.
The described effects generalize to more realistic models. This is demonstrated in simulations of a multi-laminar model of V1 composed of
leaky-integrate-and-fire (LIF) model neurons [3], where the ground
truth regarding the sub-circuits generating the oscillations is known
[4]. In this structured model these effects combine and yield misleading results. By extending mean-field theoretical descriptions of population dynamics [5] by oscillatory input, we can close the loop to the
condensed models.
We identify three main complications: First, the input can modify the
excitability of the population in a linear or non-linear fashion, yielding
significantly different changes in the spectra. Second, depending on
the properties of the system, the input to the populations is potentially low pass filtered before it enters the system. Since this low pass
filter is reflected in the response spectra, without revealing information regarding the internal dynamics of the network, we propose a
stimulation protocol counteracting this effect by emphasizing high
frequencies. Third, in general, the stimulation of a single population
excites a mixture of dynamical modes. One frequency is generated
by one dynamical mode that can be mapped to its anatomical origin
[4]. Since the observable response is composed of an inseparable mixture of modes, the mode generating the oscillation cannot easily be
isolated. Hence reconstructing the underlying connectivity as well as
identifying the role of the stimulated population in the generation of
the rhythm is not straightforward. Instead, the stimulus vector needs
to reflect the structure of the circuit generating the oscillation in order
to allow insights into the dynamically relevant components of the
system.
These problems can be regarded as a sub-set of challenges that need
to be faced when interpreting the results of circuits composed of more
complex units. The proposed solutions may be used to construct new
experimental stimulation protocols.
Acknowledgements: We acknowledge funding by the Helmholtz
Association: portfolio theme SMHB and Young Investigator’s Group
VH-NG-1028, and 604102 (Human Brain Project). All network simulations were carried out with NEST (http://www.nest-initiative.org).
References
1. Cardin JA, Carlé M, Meletis K, Knoblich U, Zhang F, Deisseroth K, Tsai L-H,
Moore CI. Driving fast-spiking cells induces gamma rhythm and controls
sensory responses. Nature. 2009;459:663–7.
2. Buzsáki G, Wang XJ. Mechanisms of gamma oscillations. Annu Rev Neurosci. 2012;35:203–25.
3. Potjans TC, Diesmann M. The cell-type specific cortical microcircuit:
relating structure and activity in a full-scale spiking network model. Cereb
Cortex. 2014;24:785–806.
4. Bos H, Diesmann M, Helias M. Identifying anatomical origins of
coexisting oscillations in the cortical microcircuit. 2015, arXiv preprint
arXiv:1510.00642.
Page 85 of 112
5.
Brunel N. Dynamics of sparsely connected networks of excitatory and
inhibitory spiking neurons. J Comput Neurosci. 2000;8:183–208.
P148
Spatiotemporal mapping of brain network dynamics
during cognitive tasks using magnetoencephalography and deep
learning
Charles M. Welzig1, Zachary J. Harper1,2
1
Departments of Neurology and Physiology, Medical College
of Wisconsin, Milwaukee, WI 53226, USA; 2College of Engineering
and Applied Science, University of Wisconsin‑Milwaukee, Milwaukee, WI
53211, USA
Correspondence: Charles M. Welzig ‑ welzig@mcw.edu
BMC Neuroscience 2016, 17(Suppl 1):P148
Magnetoencephalography (MEG) offers the high spatiotemporal resolution necessary to capture dynamic mesoscale cortical activation features
[1] for spatiotemporal mapping of brain networks. In order to adapt
such data for effective pathological or cognitive state classifiers, novel
techniques are required to extract complex connectivity dynamics that
vary in duration and latency. We have developed an advanced deeplearning system to explore such network dynamics through parcellated
connectomes in individual subjects. The machine learning system produces transparent classifiers that can define spatiotemporal characteristics of state-specific connectivity, model neurophysiological pathology
and expand understanding of connectivity dynamics. Our implementation uses source localized activation patterns extracted from event
related epochs to classify cognitive states corresponding to working
memory tasks. Subject-wise MEG data is mapped to segmented morphology from magnetic resonance imaging for source localization,
then preprocessed to optimize neural network performance. The following steps minimize dimensionality and accommodate the deeplearning system’s input requirements. First, spatiotemporal data are
encoded using a wavelet transformation that extracts oscillatory data
in the theta, alpha and beta/gamma bands per parcel (Fig. 78A). Next,
synchronicity between parcels is calculated to populate 2D connectivity matrices respective to the frequency bands. These matrices are normalized and combined into frame images where theta, alpha and beta/
gamma synchronicity is encoded as blue, green and red intensity values
respectively. These pixel grids are smoothed and expanded using gridded, cubic interpolation (Fig. 78B). The deep learning system consists of
a recursive neural network utilizing a long-short term memory (LSTM)
architecture [2] that preserves temporal input characteristics (Fig. 78C).
LSTM presents dynamically changing oscillatory patterns to the deep
learning system by integrating features of a specified range of contiguous frames relative to each training frame. As this classification system
allows for visualization of activation at each layer, we are able to identify
specific patterns that mediate the classification process (Fig. 78D) [3].
Our classification methods have demonstrated significantly low error
rate of 0.236 ± 0.425 (mean ± SD) in binary working memory state
classification after 3 min of GPU-accelerated training. Additionally,
weight patterns at specific layers within the deep learning network
Fig. 78 Visualisations of processing and deep learning stages. A
Wavelet decomposition across bands of interest. B Progression of
image-encoded oscillatory synchronization in BA10. C Deep learning
network improvement across training epochs. D One trained network
layer displaying parcel dynamics that mediate classification
BMC Neurosci 2016, 17(Suppl 1):54
highlighted relevant parcel interactions with significant effect on functional connectivity dynamics within classified cognitive states. This
project represents an advancement in preserving critical spatiotemporal information required to classify complex cognitive states that characterize dynamically changing oscillatory and synchronous functional
activity patterns across the connectome.
References
1. Larson-Prior LJ, Oostenveld R, Della Penna S, Michalareas G, Prior F,
Babajani-Feremi A, Schoffelen JM et al. Adding dynamics to the human
connectome project with MEG. NeuroImage. 2013;80:190–201.
2. Donahue J, Hendricks LA, Guadarrama S, Rohrbach M, Venugopalan S,
Saenko K, Darrell T. Long-term recurrent convolutional networks for visual
recognition and description; 2014. arXiv:14114389.
3. Plis SM, Hjelm D, Salakhutdinov R, Allen EA, Bockholt HJ, Long JD, Johnson HJ, Paulsen J, Turner JA, Calhoun VD. Deep learning for neuroimaging:
a validation study. Front Neurosci. 2014;8.
P149
Multiscale complexity analysis for the segmentation of MRI
images
Won Sup Kim1, In‑Seob Shin1, Hyeon‑Man Baek2, Seung Kee Han1
1
Department of Physics, Chungbuk National University, Cheongju,
Chungbuk 28644, Republic of Korea; 2Korea Basic Science Institute,
Cheongju, Chungbuk 28119, Republic of Korea
Correspondence: Seung Kee Han ‑ skhan@chungbuk.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P149
Segmentation of human brain images into regions with homogeneous intensity or texture is very crucial for the diagnosis of various
brain diseases. However, the presence of noises or artifacts remains
as one of the biggest obstacles for the successful segmentation. Here
we propose a novel method of segmentation based on the multiscale
complexity analysis. The idea is to characterize the complexity in visual
images by the multiscale profile representing the scale dependence
of compositional complexity. Our claim is that the multiscale profile
of human brain images combining scale dependent information on
intensity and texture information could be effectively utilized for the
segmentation of human brain images.
We have applied the multiscale complexity analysis for the segmentation of two dimensional MRI images. Our method consists of three
steps. (I) An MRI image is partitioned into homogeneous regions utilizing the information bottleneck method. (II) Multiscale complexity
profiles of individual pixels are computed from the partitioned image
of the MRI. (III) Feature vectors combining both intensity and texture
information are extracted for the segmentation. For the segmentation,
the feature vectors of individual pixels are clustered using a simple
K-mean clustering algorithm.
Using the simulated MRI images provided the BrainWeb database [1],
the performance of the segmentation was tested. The performance
shown in Fig. 79 indicates that the multiscale complexity analysis
is very robust against noise. Details will be presented during the
meeting.
Reference
1. Cocosco C, Kollokian V, Kwan RS, Evan A. BrainWeb: on line interface to a
3D MRI simulated brain database. NeuroImage. 1997;5:S425.
Fig. 79 A An MRI image from the BrainWeb [1] and the result of segmentation into five clusters. B An MRI image with 7 % noise added
and the result of segmentation into five clusters
Page 86 of 112
P150
A neuro‑computational model of emotional attention
René Richter1, Julien Vitay1, Frederick Beuth1, Fred H. Hamker1,2
1
Department of Computer Science, Chemnitz University of Technology,
Chemnitz, Germany; 2Bernstein Center for Computational Neuroscience,
Charité University Medicine, Berlin, Germany
Correspondence: René Richter ‑ rene.richter@cs.tu‑chemnitz.de
BMC Neuroscience 2016, 17(Suppl 1):P150
Emotional stimuli attract attention so the brain can focus its processing resources on them. The questions that arise is how these stimuli
acquire their emotional value and how they can influence attentional
processes. Evidence suggests that this association might be learned
through conditioning in the amygdala, more specifically the basal
lateral amygdala (BLA). Furthermore, feedback connections from the
BLA to the visual cortex seem to enhance the activation of neural representations which is a possible top-down attention mechanism of
the emergent attention hypothesis. While neuro-computational models of attention mechanisms attract increasing interest due to their
importance for the focused processing of information in the brain, the
possible emotional feedback from the amygdala is to date largely unexplored. Therefore, we propose a rate-coded, biological realistic neurocomputational model constructed of 3 smaller functional models. First,
we combined a model of the visual processing pathway for object recognition [1] that includes the retina, the lateral geniculate nucleus, the
visual areas V1, V2 and V4 as well as the frontal eye field with an amygdala model for the associative conditioning of a visual stimulus with a
bodily reaction representing a particular emotional state. Second, in
order to provide the model with realistic temporal learning properties,
a reward-timing model [2] simulating the afferent system to the dopaminergic area VTA has been integrated to temporally adjust the learning process through dopamine-mediated modulation of plasticity. This
timing model includes a number of brain areas, most prominently the
ventral tegmental area, the nucleus accumbens, the lateral hypothalamus, the ventral medial prefrontal cortex and the amygdala. In order
to enable emotional attention, 2 simulation phases were implemented:
(1) a conditioning phase to learn the association between an important stimulus and the body reaction, and (2) an attention phase where
the representation of the visual stimulus activates the BLA which then
sends back a feedback to enhanced this specific stimulus. Afterwards,
the enhanced representation in V4 suppresses the competing ones
and allows the frontal eye field to initiate a saccade in its direction. As
a result of the biologically based connectivity and the realistic learning
process, the model outcomes are coherent with several experimental
findings and increase our understanding of the brain network’s interaction. In the future, the model could furthermore be used for facial
analysis and the process of learning the importance of specific facial
features for emotional expressions.
References
1. Beuth F, Hamker FH. A mechanistic cortical microcircuit of attention for
amplification, normalization and suppression. Vis Res. 2015;116:241–57.
2. Vitay J, Hamker FH. Timing and expectation of reward: a neuro-computational model of the afferents to the ventral tegmental area. Front
Neurorobot. 2014;8:1–25.
P151
Multi‑site delayed feedback stimulation in parkinsonian networks
Kelly Toppin1, Yixin Guo1
1
Department of Mathematics, Drexel University, Philadelphia, PA 19104,
USA
Correspondence: Yixin Guo ‑ yixin@math.drexel.edu
BMC Neuroscience 2016, 17(Suppl 1):P151
The conventional deep brain stimulation (DBS), as a surgical procedure
to alleviate debilitating and disrupting symptoms of Parkinson’s disease (PD), has several drawbacks. Multi-site delayed feedback stimulation (MDFS) has been proposed as a feasible alternative to overcome
the drawbacks of the conventional DBS [2, 3]. We first build two types
of large scale biophysical networks to explore the effectiveness of
MDFS. The persistent parkinsonian network has strongly synchronized
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 80 A The network model. Plus symbol indicates excitatory
connection, and minus symbol indicates inhibitory connection. LFP
is computed from GPi and STN populations separately. These two
LFP signals are used as the source of the five MDFS stimulations
(dashed-arrows): STN-to-STN, STN-to-GPi, GPi-to-STN, GPi-to-GPe and
GPi-to-GPi, shown by arrow labeled 1–5. B Error index values for 80
different model TC neurons in an intermittent network. Comparison
of MDFS among different stimulation targets using either STN or GPi
LFP signal. Whisker plots show mean (red line), 25–75 percentile range
(blue box), 95 % confidence interval (black lines) and outliers (red plus
signs)
bursting clusters with elevated firing rates present in subthalmic
nucleus (STN), internal and external segments of globus pallidus (GPi
and GPe) neurons. However, the brain of a PD patient may not be in
a constant strong synchronized and clustered state, and short desynchronized events may present when the brain is in between high synchronization [1]. We build an intermittent parkinsonian network that
can transit between synchronized and desynchronized dynamics.
Using both parkinsonian networks, we compute the TC error index, the
fraction of miss responses and excessive responses when a TC neuron
relays multiple excitatory inputs, in five different stimulation settings:
MDFS from STN to STN, from STN to GPe, from GPi to STN, from GPi to
GPe and from GPi to GPi, shown by the dashed-arrows labeled 1–5 in
Fig. 80A. Each “to” population is stimulated by the signal based on the
LFP calculated at the “from” population. Our results of lower TC relay
errors with the five different stimulations in Fig. 80B show that MDFS
improves the fidelity of the TC relay neurons’ communication and
responses to the input motor signal in both persistent and intermittent parkinsonian networks. We also find that MDFS with STN or GPe
as a stimulation target is more effective in reducing TC relay errors.
Acknowledgements: The study was supported by NSF grant DMS1226180 awarded to Yixin Guo.
References
1. Ahn S, Rubchinsky L. Short desynchronization episodes prevail in synchronous dynamics of human brain rhythms. Chaos. 2013;23:013138.
2. Guo Y, Rubin JE. Multi-site stimulation of subthalamic nucleus diminishes
thalamocortical relay errors in a biophysical network model. Neural Netw.
2011;24(6):602-16.
3. Hauptmann C, Omel’Chenko O, Popovych V, Maistrenko Y, Tass PA.
Control of spatially patterned synchrony with multisite delayed feedback.
Phys Rev E. 2007;76(6):066209.
P152
Bistability in Hodgkin–Huxley‑type equations
Tatiana Kameneva1, Hamish Meffin2, Anthony N. Burkitt1, David B
Grayden1,3
1
NeuroEngineering Laboratory, Department of Electrical & Electronic
Engineering, University of Melbourne, Parkville, VIC 3010, Australia;
2
National Vision Research Institute, Australian College of Optometry,
Carlton, VIC 3053, Australia; 3Centre for Neural Engineering, University
of Melbourne, Parkville, VIC 3010, Australia
Correspondence: René Richter ‑ tkam@unimelb.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P152
Page 87 of 112
Fig. 81 Response of a modeled neuron for different initial conditions, V(0). Sodium, three types of potassium, calcium, and hyperpolarisation-activated currents are included in the model
Background Purkinje cells have two states of the resting membrane
potential: a hyperpolarized quiescent state (down state) and a depolarized spiking state (up state) [1]. This bistability has been observed
in in vitro and in vivo recordings, in anesthetized animals, and in slices.
It has been proposed that bistability in Purkinje cells play a key role in
the short-term processing and storage of sensory-motor information.
Methods To investigate bistability of the neuronal resting state, we
use computer simulations in neuron. We simulate single compartment
neurons and use the Hodgkin–Huxley-type formalism to study how
initial conditions and a combination of ionic channels affect neuronal
response. We systematically apply intracellular current pulse stimulation to set the membrane potential to different levels and observe the
neuronal dynamics after the stimulation is released.
Results We show that the neural response after release of the pulse stimulation depends on the amplitude of the current pulses. For some stimulation levels, the cells return to the level of the activity prior to stimulation,
while for other levels, the neuronal dynamics are different to prior activity
levels for a long time post stimulation. We show that different initial conditions lead to different neuronal dynamics even when all other parameters
in the Hodgkin–Huxley-type model are set the same (Fig. 81). We explore
the region of attraction for two stable states and find that they differ for
different parameters of the model, in particular that different ionic channel combinations do not change our qualitative results.
Conclusions This work demonstrates a potential method to explore
the mechanisms underlying bistability in Purkinje cells. In particular,
the proposed methodology allows the exploration of the circumstances under which Purkinje cells transit from the down state to the
up state and return. This work implies that results obtained using the
Hodgkin–Huxley formalism should be carefully considered since the
choice of initial conditions may significantly affect the final outcome.
Acknowledgements: This research was supported by the Australian
Research Council (ARC). TK acknowledge support through ARC Discovery Early Career Researcher Award (DE120102210).
Reference
1. Loewenstein Y, Mahon S, Chadderton P, Kitamura K, Sompolinsky H,
Yarom Y, Hausser. Bistability of cerebella Purkinje cells modulated by
sensory stimulation. Nat Neurosci. 2005;8(2):202–11.
P153
Phase changes in postsynaptic spiking due to synaptic
connectivity and short term plasticity: mathematical analysis
of frequency dependency
Mark D. McDonnell1, Bruce P. Graham2
1
Computational and Theoretical Neuroscience Laboratory, School
of Information Technology and Mathematical Sciences, University
of South Australia, Mawson Lakes, SA, 5095, Australia; 2Computing
Science and Mathematics, School of Natural Sciences, University
of Stirling, Stirling, FK9 4LA, UK
Correspondence: Mark D. McDonnell ‑ mark.mcdonnell@unisa.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P153
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 82 The phase of post-synaptic firing in response to frequencymodulated inhomogenous Poisson pre-synaptic spike trains depends
on the configuration of input synapses. A The connectivity involves
M independent pre-synaptic neurons each with NM synaptic release
sites at a post-synaptic neuron. Vesicles are released probabilistically
when activated by a pre-synaptic action-potential, if one is available
at that site. The post-synaptic neuron depolarizes and produces
action potentials after arrival of neurotransmitter according to standard models. B The phase lead of output spiking relative to periodic
modulation at frequency f, for M active-zones is both frequencydependent and configuration-dependent. The figure is from simulations but we also derived the same result mathematically
We examined how short-term synaptic depression due to vesicle
depletion [1] interacts with the configuration of the synaptic pathways
onto an output neuron. Using both simulations and mathematical
analysis, we found significant frequency-dependent phase-shifts of
the spiking response of a neuron driven by independent frequencymodulated Poisson input signals. The synaptic inputs to the neuron
are assumed to consist of a fixed number of release sites that are
divided between active zones, with each active zone being the presynaptic axonal target of a single input neuron (Fig. 82A). For the same
number of release sites, at one extreme the output neuron receives
input from a large number of neurons through independent active
zones, each containing a single release site, similar to cortical cells.
At the other extreme (similar to a Calyx of Held in the auditory brainstem), the neuron is driven by a single input neuron through a giant
synapse containing a single active zone with a very large number of
release sites.
Using standard stochastic models of short term depression due to
vesicle depletion [2], and post-synaptic current dynamics, we found
strong phase dependencies for input modulation rates up to 5 Hz.
The phase shift also depends strongly on the configuration (Fig. 82B).
However, the phase shift otherwise remains invariant for a wide range
of post-synaptic conditions, such as for Hodgkin–Huxley or leaky integrate-and-fire models, and whether or not the dynamics of post synaptic currents included rise-times, or longer or shorter decay times.
Acknowledgements: M. D. McDonnell was supported by an Australian Research Fellowship from the Australian Research Council (project
DP1093425). B. P. Graham’s contribution was supported by the BBSRC
project grant BB/K01854X/1.
References
1. Abbott LF, Regehr WG. Synaptic computation. Nature. 2004;431:796–803.
2. McDonnell MD, Mohan A, Stricker C. Mathematical analysis and algorithms for efficiently and accurately implementing stochastic simulations of short-term synaptic depression and facilitation. Front Comput
Neurosci. 2013;7:58.
P154
Quantifying resilience patterns in brain networks: the importance
of directionality
Penelope J. Kale1, Leonardo L. Gollo1
1
Systems Neuroscience Group, QIMR Berghofer Medical Research
Institute, Brisbane, QLD, 4006, Australia
Correspondence: Penelope J. Kale ‑ Penelope.Kale@qimr.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P154
Page 88 of 112
Defining how interactions take place, directionality is major feature
of network connections. Brain networks are intrinsically directed
because of the nature of chemical synapses, which comprise most of
the neuronal connections. The specific fingerprint of the interactions
between cortical regions and neurons thereof are crucial to the neuronal dynamics. The neuronal ability to synchronize is extremely sensitive to the presence of reciprocal connections in neuronal motifs and
circuits [1]. The type of synchronization (or the phase relation between
phase locked neurons and cortical regions) also depends on the relation between the synaptic strengths between these regions [2, 3].
Moreover, whole brain network dynamics is also shaped by reciprocal
connections, which stabilizes the network dynamics and reduce transitions between metastable states [4]. However, due to limitations in
current brain imaging techniques, the directionality of edges between
structurally connected regions of the human brain cannot be confirmed. Additionally, despite the demonstrated importance of synaptic
direction, its effect over main network features is not yet elucidated.
Comparing several directed brain networks from different species
(macaque, cat, mouse, and C. elegans) and with variable node size
(parcellation), we estimate the error that is made in characterizing
and identifying brains as complex network when undirected networks
are assumed. We use different approaches to turn directed networks
undirected: (i) remove unidirectional links; (ii) add reciprocal links; (iii)
add one reciprocal for each removed unidirectional link thus keeping
the same network density. We find that directionality plays a major
role in shaping the brain networks. All regions are affected, including hub nodes, which have large degree and enhanced importance
in information integration for cognitive functions [5]. We compute
and rank graph theoretical measures and determine their resilience
with respect to the loss of directionality of the network. Overall, our
results suggest that the characterization of connectomes can be compromised in the absence of data regarding the directionality of brain
networks.
References
1. Gollo LL, Mirasso C, Sporns O, Breakspear M. Mechanisms of zero-lag synchronization in cortical motifs. PLoS Comput Biol. 2014;10(4):e1003548.
2. Matias FS, Carelli PV, Mirasso CR, Copelli M. Anticipated synchronization in a biologically plausible model of neuronal motifs. Phys Rev E.
2011;84(2):021922.
3. Matias FS, Gollo LL, Carelli PV, Bressler SL, Copelli M, Mirasso CR. Modeling
positive Granger causality and negative phase lag between cortical areas.
NeuroImage. 2014;99:411–8.
4. Gollo LL, Zalesky A, Hutchison RM, van den Heuvel M, Breakspear
M. Dwelling quietly in the rich club: brain network determinants
of slow cortical fluctuations. Philos Trans R Soc Lond B Biol Sci.
2015;370(1668):20140165.
5. van den Heuvel MP, Sporns O. Network hubs in the human brain. Trends
Cogn Sci. 2013;17(12):683–96.
P155
Dynamics of rate‑model networks with separate excitatory
and inhibitory populations
Merav Stern1, L.F. Abbott2
1
Faculty of Medicine, Technion, Haifa, Israel; 2Department of Neuroscience
and Department of Physiology and Cellular Biophysics, Columbia
University, New York, NY, USA
Correspondence: Merav Stern ‑ merav.stern@mail.huji.ac.il
BMC Neuroscience 2016, 17(Suppl 1):P155
Randomly connected networks of rate-model neurons have a rich
dynamics [1], a feature that has been exploited to model a variety of
phenomena [2]. These model networks typically do not distinguish
between excitatory and inhibitory neuron classes. Doing this requires
constraining the network connectivity matrix to have columns with
exclusively positive entries, representing input from excitatory neurons, and with negative entries, representing input from inhibitory
neurons. The eigenvalue spectra of random matrices satisfying this
constraint have a number of interesting properties [3, 4].
Here we study the dynamics of rate-model networks that result
from using such connectivity matrices. We find that neural activity is
BMC Neurosci 2016, 17(Suppl 1):54
correlated across all neurons, including both excitatory and inhibitory
subpopulations. This correlation depends on the difference between
the mean strengths of the excitation and inhibition connections and
it increases as this difference is increased. For very large values of this
difference, the network reaches a stable fixed point, otherwise it is
chaotic. Chaos arises from the residual activity deviating from the correlated mean network activity and it acts to reduce these correlations.
The magnitude of the residual chaotic activity is determined by the
variances of the synaptic strengths within the excitatory and inhibitory
populations.
In summary, unlike models with a single mixed excitatory/inhibitory
population, in which the activity between pairs of neurons is uncorrelated for every value of synaptic gain, networks with distinct excitatory and inhibitory subpopulations exhibit strongly correlated activity
across the entire network reminiscent of the up/down states seen in
neural recordings [5].
References
1. Sompolinsky H, Crissanti A, Sommers HJ. Chaos in random neural networks. Cerebral Phys Rev Lett. 1988;61:259–62.
2. Sussillo D. Neural circuits as computational dynamical systems. Curr Opin
Neurobiol. 2014;25:156–63.
3. Rajan K, Abbott LF. Eigenvalue spectra of random matrices for neural
networks. Phys Rev Lett. 2006;97:188104.
4. Tao T. Outliers in the spectrum of IID matrices with bounded rank perturbations. arXive: 1012.4818v6.
5. Steriade M, Nunez A, Amzica F. A novel slow (<1 Hz) oscillation of neocortical neurons in vivo: depolarizing and hyperpolarizing components. J.
Neurosci. 1993;13:3252–65.
P156
A model for multi‑stable dynamics in action recognition
modulated by integration of silhouette and shading cues
Leonid A. Fedorov1,2, Martin A Giese1,2
1
Section for Computational Sensomotorics, Department of Cognitive
Neurology, CIN&HIH, Tübingen, Germany; 2GTC, International Max Planck
Research School, University of Tübingen, Tübingen, Germany
Correspondence: Leonid A. Fedorov ‑ leonid.fedorov@uni‑tuebingen.de
BMC Neuroscience 2016, 17(Suppl 1):P156
The visual perception of body motion can show interesting multi-stability. For example, a walking body silhouette (bottom inset Fig. 83A)
is seen alternately as walking in two different directions [1]. For stimuli
with minimal texture information, such as shading, this multi-stability
disappears. Existing neural models for body motion perception [2–4]
do not reproduce perceptual switching. Extending the model [2], we
developed a neurodynamic model that accounts for this multi-stability
(Fig. 83A). The core of the model is a two-dimensional neural field that
consists of recurrently coupled neurons with selectivity for instantaneous body postures (‘snapshots’). The dimensions of the field encode
Fig. 83 A Model architecture with 2D neural field that receives input
from two hierarchical path-ways. B Response traces of MP neurons
for silhouette stimulus without shading during a 200 s simulation.
C Corresponding average response times of the output neurons. D
Response times for shaded stimulus
Page 89 of 112
the keyframe number θ and the view of the walker ф. The lateral connectivity of the field stabilizes two competing traveling pulse solutions that encode the perceived temporally changing action patterns
(walking in the directions ±45°). The input activity of the field is generated by two visual pathways that recognize body postures from graylevel input movies. One pathway (‘silhouette pathway’) was adapted
from [2] and recognizes shapes, mainly based on the contrast edges
between the moving figure and the background. The second pathway
is specialized for the analysis of luminance gradients inside the moving figure. Both pathways are hierarchical (deep) architectures, built
from detectors that reproduce known properties of cortical neurons.
Higher levels of the hierarchies extract more complex features with
higher degree of position/scale invariance. The field activity is read out
by two Motion Pattern (MP) neurons, which encode the two possible
perceived walking directions. Testing the model with an unshaded silhouette stimulus, it produces randomly switching percepts that alternate between the walking directions (±45°) (Fig. 83B, C). Addition of
shading cues disambiguates the percept and removes the bistability
(Fig. 83D). The developed architecture accounts for the disambiguation by shape-from shading.
Acknowledgements: Supported by EC Fp7-PEOPLE-2011-ITN PITNGA-011-290011 (ABC), FP7-ICT-2013-FET-F/604102 (HBP), FP7-ICT-201310/611909 (Koroibot), BMBF, FKZ: 01GQ1002A, DFG GI 305/4-1 + KA
1258/15-1.
References
1. Vangeneugden J, et al. Activity in areas MT+ and EBA, but not pSTS,
allows prediction of perceptual states during ambiguous biological
motion. Soc Neurosci Meet. 2012;127(04).
2. Giese MA, Poggio T. Neural mechanisms for the recognition of biological
movements and action. Nat Rev Neurosci. 2003;4:179–92.
3. Lange J, Lappe M. A model of biological motion perception from configural form cues. J Neurosci. 2006;26:2894–906.
4. Jhuang H, et al. A biologically inspired system for action recognition. In:
ICCV 2007. p. 1–8.
P157
Spiking model for the interaction between action recognition
and action execution
Mohammad Hovaidi Ardestani1,2, Martin Giese1
1
Section Computational Sensomotorics, CIN & HIH, Department
of Cognitive Neurology, Tübingen, 72076, Germany; 2IMPRS for Cognitive
and Systems Neuroscience, University Clinic Tübingen, Tübingen, 72076,
Germany
Correspondence: Mohammad Hovaidi Ardestani ‑ Mohammad.
Hovaidi‑Ardestani@uni‑tuebingen.de
BMC Neuroscience 2016, 17(Suppl 1):P157
Action perception and the control of action execution are intrinsically linked in the human brain. Experiments show that the concurrent motor execution influences the visual perception of actions
and biological motion (e.g. [1]). This interaction likely is mediated by
action-selective neurons in the STS, premotor and parietal cortex.
We have developed a model based on biophysically realistic spiking
neurons that accounts for the observed interactions between action
perception and motor planning. The model is based on two dynamic
representation levels (Fig. 84A), one modeling a representation of
perceived action patters (vision field), and one representing associated motor programs (motor field). Both levels are modeled by recurrent spiking networks that approximate neural fields, where each field
consists of 30 coupled neural ensembles, each consisting of 80 excitatory and 20 inhibitory adaptive Exponential Integrate-and-Fire (aEIF)
neurons [2]. Within each field asymmetric recurrent connections
between the ensembles stabilize a traveling pulse solution, which is
stimulus-driven in the visual field and autonomously propagating in
the motor field after initiation by a go-signal. Both fields are coupled
by interaction kernels that results in mutual excitation between the
fields of the traveling pulse propagate synchronously and in mutual
inhibition otherwise. We used the model to reproduce the result of a
psychophysical experiment that tested the detection of point-light
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 84 A Model architecture consisting of two coupled neural fields,
implemented with biophysically realistic neurons. B Psychophysical
results from [1] showing the dependence of the detectability of visual
point-light stimuli in dependence of the delay between a visually
observed and the concurrently executed action. C Simulated detectability derived from the model for the same experimental conditions
stimuli in noise during concurrent motor execution [1]. The pointlight patterns showed arm movements of the observer, which were
synchronized with varying time delays with the executed movements. Compared to a baseline without concurrent motor execution,
the detectability of the visual stimulus was higher for very small time
delays between the visual stimulus and the executed arm movement, and it was lower when the observed movement was strongly
delayed (>300 ms) against the executed motor patterns (Fig. 84B).
The same pattern arises from the detectability of the visual stimulus
as predicted from our model, where we assumed that the level of
neural activity (compared to a noise level) provides a measure for the
detectability of the stimulus (Fig. 84C). The proposed model, which is
derived by simplification from physiologically-inspired neural models
for action execution and motor planning, reproduces correctly the
modulation of visual detection by the synchrony of the stimulus with
executed motor behavior. Present work extends the model by a full
visual pathway and an effector model, allowing for the simulation of a
broader spectrum of experimental results.
Acknowledgements: Supported by EC FP7-ICT-2013-FET-F/604102
(HBP), Fp7-PEOPLE-2011-ITN PITN-GA-011-290011 (ABC), FP7ICT-2013-10/611909 (Koroibot), BMBF, FKZ: 01GQ1002A, DFG GI
305/4-1 + KA 1258/15-1.
References
1. Christensen A, Ilg W, Giese MA. Spatiotemporal tuning of the facilitation
of biological motion perception by concurrent motor execution. J Neurosci. 2011;31:3493–9.
2. Brette R, Gerstner W. Adaptive exponential integrate-and-fire model
as an effective description of neuronal activity. J Neurophysiol.
2005;94:3637–42.
P158
Surprise‑modulated belief update: how to learn within changing
environments?
Mohammad Javad Faraji1, Kerstin Preuschoff2, Wulfram Gerstner1
1
School of Life Sciences, Brain Mind Institute and School of Computer
and Communication Sciences, Ecole Polytechnique Federal de Lausanne
(EPFL), CH‑1015 Lausanne, Switzerland; 2Geneva Finance Research
Institute (GFRI) and Swiss Center for Affective Sciences (CISA), University
of Geneva, CH‑1211 Geneva, Switzerland
Correspondence: Mohammad Javad Faraji ‑ mohammadjavad.faraji@
epfl.ch
BMC Neuroscience 2016, 17(Suppl 1):P158
Page 90 of 112
Surprise is informative because it drives attention [1] and modifies
learning [2]. Correlates of surprise have been observed at different
stages of neural processing, and found to be relevant for learning
and memory formation [3]. Although surprise is ubiquitous, there is
neither a widely accepted theory that quantitatively links surprise to
observed behavior, such as the startle response, nor an agreement on
how surprise should influence learning speed or other parameters in
iterative statistical learning algorithms. Building on and going beyond
earlier surprise measures [4–6], we propose a novel information theoretic measure for calculating surprise in a Bayesian framework so as
to capture uncertainty of the world as well as imperfections of the
subjective model of the world, two important aspects of surprise.
The principle of future surprise minimization leads to a learning rule
that can be interpreted as a surprise modulated belief update suitable for learning within changing environments. Importantly, we do
not need an assumption on how quickly the world changes. We apply
our surprise-modulated learning rule to an exploration task in a mazelike environment. Our results are consistent with the behavioral finding that surprising events induce humans and animals to learn faster
and to adapt more quickly to changing environments. Information
content [5] captures the inherent unexpectedness of a piece of data
for a given set of models (uncertainty of the world), whereas Bayesian
surprise [4] measures the change in belief caused by a new data point
(observer dependent). These are two complementary approaches for
calculating surprise. In our approach both aspects are combined with
a third aspect: if we are uncertain about what to expect, receiving a
low-probability data sample is less surprising than in a situation when
we are almost certain about the world. A surprise minimizing learning
(SMiLe) rule is derived by solving a constrained optimization problem defined as follows: the objective is to maximally reduce surprise
when facing the same data again in the not so far future, under the
constraint that the posterior belief (after the update step) is not too
different from the prior. The resulting SMiLe rule balances the influence of newly acquired data with prior knowledge where the balance
depends on surprise. In case of a fundamental change in the world
signaled by surprising samples, data acquired before the change is
downgraded as less informative about the current state of the world.
A simultaneous increase of the influence of newly acquired data on
learning leads to a fast adaptation of the model to an environmental
change. While in a stationary environment our algorithm approaches
the known Bayesian update rule, it also allows the model to react to
changes in the environment. In summary, surprising data increases the
uncertainty we have about our current model of the world and gives
a bigger influence of newly acquired data on belief update. The interaction between surprise and uncertainty is important for modeling
the behavior of humans and animals in changing environments. The
surprise signal could be broadly transmitted in the brain by a neuromodulator with widespread axonal ramifications (e.g., norepinephrine
(NE) released from locus coeruleus (LC) neurons) and influence synaptic plasticity rules.
Acknowledgements: This research was supported by the European
Research Council (Grant Agreement No. 268 689).
References
1. Itti L, Baldi P. Bayesian surprise attracts human attention. Vis Res.
2009;49(10):1295–1306.
2. Schultz W, Dickinson A. Neuronal coding of prediction errors. Annu Rev
Neurosci. 2000;23(1):473–500.
3. Ranganath C, Rainer G. Neural mechanisms for detecting and remembering novel events. Nat Rev Neurosci. 2003;4(3):193–202.
4. Baldi P, Itti L. Of bits and wows: a bayesian theory of surprise with applications to attention. Neural Netw. 2010;23(5):649–66.
5. Shannon CE. A mathematical theory of communication. ACM SIGMOBILE
Mobile Comput Commun Rev. 2001;5(1):3–55.
6. Palm G. Novelty, information and surprise. Berlin: Springer; 2012.
BMC Neurosci 2016, 17(Suppl 1):54
P159
A fast, stochastic and adaptive model of auditory nerve responses
to cochlear implant stimulation
Margriet J. van Gendt1, Jeroen J. Briaire1, Randy K. Kalkman1, Johan H. M.
Frijns1,2
1
ENT‑Department, Leiden University Medical Centre, Leiden, 2300 RC, the
Netherlands; 2Leiden Institute for Brain and Cognition, Leiden, 2300 RC,
the Netherlands
Correspondence: Margriet J. van Gendt ‑ m.j.van_gendt@lumc.nl
BMC Neuroscience 2016, 17(Suppl 1):P7
Cochlear implants (CI) rehabilitate hearing impairment through direct
electrical stimulation of the auditory nerve. In many modern CIs sound
is coded through the continuous interleaved sampling (CIS) strategy.
Although many different sound-coding strategies have been introduced in the last decade, no major advances have been made since
the introduction of the CIS strategy [1]. New stimulation strategies are
commonly investigated by means of psychophysical experiments and
clinical trials, which is time-consuming for both patient and researcher.
Alternatively, strategies can be evaluated using computational models. In this study a computationally efficient model that accurately predicts auditory nerve responses to CI pulse train input is developed.
The model includes the 3D volume conduction and active nerve model
developed in the Leiden University Medical Center [2], and is extended
with stochasticity, adaptation and accommodation. This complete
model includes spatial as well as temporal characteristics of both the
cochlea and the auditory nerve. The stochastic and adaptive auditory
nerve model is used to investigate full-nerve responses to amplitude
modulated long duration stimulation. Understanding responses to
amplitude modulation is important because current speech coding
strategies are based on the principle of speech information distribution
through amplitude modulation of the input pulse trains. The model
is validated by comparison to experimentally measured single fiber
action potential (SFAP) responses to pulse trains published in literature
[3–6]. The effects of different pulse-train parameters such as pulse rate,
pulse amplitude and amplitude modulation are investigated.
The neural spike patterns produced in response to CI stimulation are
very similar to spike patterns obtained with single fiber action potential measurements in animal experiments in response to CI stimulation. Besides predicting single fiber responses to constant amplitude
pulse trains, the model also very well predicts single fiber responses
to amplitude modulated pulse trains. Response alterations seen over
the duration of the stimulus are similar to those seen in animal experiments. Modeled effects of stimulus amplitude, pulse rate and amplitude modulation is similar to the effects seen in animal experiments.
Adaptation is found to be an important factor in modeling nerve outcomes to amplitude modulated pulse trains and their spatial effects.
The model is shown to accurately predict spike timings in response
to long duration pulse trains as observed in animal experiments. The
model can be used to predict full auditory nerve responses to electrical pulse trains, and thus to different sound coding strategies. The next
step will be to apply this model to evaluate complete auditory nerve
responses to different sound coding strategies.
Acknowledgements: This study was financially supported by
Advanced Bionics Corporation.
References
1. Zeng FG, Rebscher S, Harrison WV. Cochlear implants: system design,
integration and evaluation. IEEE Rev Biomed Eng. 2008;115–42.
2. Kalkman RK, Briaire JJ, Dekker DMT, Frijns JHM. Place pitch versus
electrode location in a realistic computational model of the implanted
human cochlea. Hear Res. 2014;315:10–24.
3. Miller CA, Hu N, Zhang F, Robinson BK, Abbas PJ. Changes across time in
the temporal responses of auditory nerve fibers stimulated by electric
pulse trains. J Assoc Res Otolaryngol. 2008;9:122–37.
4. Litvak L, Delgutte B, Eddington D. Auditory nerve fiber responses to
electric stimulation: modulated and unmodulated pulse trains. J Acoust
Soc Am. 2001;110:368.
Page 91 of 112
5.
6.
Zhang F, Miller CA, Robinson BK, Abbas PJ, Hu N. Changes across time
in spike rate and spike amplitude of auditory nerve fibers stimulated by
electric pulse trains. JARO J Assoc Res Otolaryngol. 2007;8:356–72.
Hu N, Miller CA, Abbas PJ, Robinson BK, Woo J. Changes in auditory nerve
responses across the duration of sinusoidally amplitude-modulated
electric pulse-train stimuli. J Assoc Res Otolaryngol. 2010;11:641–56.
P160
Quantitative comparison of graph theoretical measures
of simulated and empirical functional brain networks
Won Hee Lee1, Sophia Frangou1
Department of Psychiatry, Icahn School of Medicine at Mount Sinai, New
York, NY 10029, USA
Correspondence: Won Hee Lee ‑ wonhee.lee@mssm.edu
BMC Neuroscience 2016, 17(Suppl 1):P160
Graph theoretical approaches to resting-state fMRI have been widely
used to quantitatively characterize functional network organization
in the resting brain, but mechanistic explanations for how restingstate brain works are still lacking. Whole-brain computational models
have shown promise in enriching our understanding of mechanisms
contributing to the formation and dissolution of resting-state functional patterns [1]. It is therefore important to determine the degree
to which computational models reproduce the topological features
of empirical functional brain networks. Here, we focused on the performance of the Kuramoto model [2] as it is considered most representative model of coupled phase oscillators and is widely used in
the literature.
Empirical and simulated functional networks were defined based on
66 brain anatomical regions (nodes). Simulated resting-state functional connectivity (FC) was generated using the Kuramoto model
constrained by empirical structural connectivity. The simulated FC
matrix was tuned to best fit empirical FC matrix. In order to improve
stability and reliability, we simulated 10 runs of fMRI BOLD time
series (obtained from 320 s simulations, discarding 20 s initial transients) with varying random initial conditions, and generated the
best-fit simulated FC matrix for each run. We applied graph theoretical approaches to optimally simulated FC and empirical FC data to
characterize key topological features of brain networks [3]. Finally,
we quantified and compared the difference, in terms of relative error,
in graph theoretical measures between the simulated and empirical
functional networks.
Figure 85 shows the quantitative difference in graph theoretical measures between the empirical FC and the simulated FC over the entire
(1–100 %) and selected range of connection densities (37–50 %).
The averaged relative differences were found to be 2–77 % over the
entire range of connection densities as well as 0.1–22 % over a range
of 37–50 % connection densities. We found that simulated functional
data can be used with confidence to model graph measures of global
and local efficiency, characteristic path length, eigenvector centrality,
and resilience to targeted attack and random failure. Our results also
highlight the critical dependence of the solutions obtained in simulated data on the specified connection density.
This study demonstrates the value of computational models in assessing whole-brain network connectivity, and provides a method for the
quantitative evaluation and external validation of graph theory metrics derived from simulated data that can be used to inform future
study designs.
References
1. Deco G, Jirsa VK, McIntosh AR: Emerging concepts for the dynamical
organization of resting-state activity in the brain. Nat Rev Neurosci.
2011;12(1):43–56.
2. Kuramoto Y. Chemical oscillations, waves, and turbulence. Berlin:
Springer; 1984.
3. Rubinov M, Sporns O. Complex network measures of brain connectivity:
uses and interpretations. Neuroimage. 2010l52:1059–69.
BMC Neurosci 2016, 17(Suppl 1):54
Page 92 of 112
identified over 100 time-series features of the BOLD signal with statistically significant separability between SCZ and CON (p < 0.05). These
features were mostly measures of time series ‘predictability’, including
autocorrelation, local prediction error (using exponential smoothing,
Gaussian Processes, and AR models), the SD1 measure from the heart
rate variability literature, and low frequency power. This emergent
class of discriminative properties of BOLD dynamics is consistent with
the use of the ALFF metric in existing work using fMRI data [3].
We present the first systematic comparison of thousands of interdisciplinary time-series analysis measures to fMRI data and use machine
learning to uncover characteristic BOLD signatures of schizophrenia,
in both space and time. In a completely data-driven manner, we identify informative brain regions and time-series analysis techniques that
best discriminate people with schizophrenia from healthy controls,
using just the properties of BOLD signals in individual ROIs. The framework presented here represents a general and powerful data-driven
means of identifying discriminative time-series features from neuroscience data.
Fig. 85 Relative error (RE) in percentage between graph theoretical
measures of simulated FC versus empirical FC for the entire (1–100 %)
and selected range of connection densities (37–50 %). Bars and error
bars correspond respectively to the averages and standard deviations
across the ten RE values. Eglob global efficiency, Eloc local efficiency,
CC clustering coefficient, L characteristic path length, EC eigenvector
g
centrality, PC participation coefficient, SW: small-worldness, Rct and Rt
represent resilience to targeted attack in the size of largest connected
g
component and global efficiency, respectively, Rcr and Rr represent
resilience to random failure in the size of largest connected component and global efficiency, respectively
P161
Determining discriminative properties of fMRI signals
in schizophrenia using highly comparative time‑series analysis
Ben D. Fulcher1, Patricia H. P. Tran1, Alex Fornito1
1
Monash Institute of Cognitive and Clinical Neurosciences, Monash
University, Clayton, Vic 3168, Australia
Correspondence: Ben D. Fulcher ‑ ben.fulcher@monash.edu
BMC Neuroscience 2016, 17(Suppl 1):P161
Analysis of fMRI data typically focuses on inter-regional functional connectivity, measured as pairwise correlations, or through multivariate
decompositions (e.g., ICA). Relatively little attention is given to the univariate time-series properties of BOLD signals within a specific brain
region, despite a broad scientific literature on time-series analysis
(including power spectral techniques, information theoretic methods,
model fitting, nonlinear time-series analysis, and fractal scaling). Here
we undertake the largest systematic comparison of over 7000 such
measures of temporal structure to identify the temporal features of
individual BOLD signals, and their locations in the brain, that are most
discriminative of people with schizophrenia.
MRI data were obtained from the open COBRE database [4] for 72 people with schizophrenia (SCZ) and 74 healthy controls (CON). For each
subject, we extracted 7779 temporal features from the BOLD time
series recorded in each of 264 brain regions using the publicly available highly comparative time-series analysis framework, hctsa (http://
benfulcher.github.io/hctsa/) [1].
Spatial analysis ROIs that were most discriminative of SCZ versus CON
were identified by training a separate linear support vector machine
(SVM) classifier for each ROI, across all features, using tenfold cross validation. We identified 23 ROIs with a classification accuracy exceeding
chance levels (p < 0.05, FDR-corrected) with some individual ROI accuracies reaching 69.5 %. These discriminative brain regions were mostly
located in the frontal and parietal cortices.
Temporal analysis The most discriminative temporal features were
deduced using t-tests in each ROI, and then averaging across all ROIs. P
values were computed using permutation tests with 1000 shuffles. We
References
1. Fulcher BD, Little MA, Jones NS. Highly comparative time-series analysis:
the empirical structure of time series and their methods. J R Soc Interface.
2013;10:20130048.
2. Power JD, Cohen AL, Nelson SM, Wig GS, Barnes KA, Church JA, Vogel AC,
Laumann TO, Miezin FM, Schlaggar BL, Petersen SE. Functional network
organization of the human brain. Neuron. 2011;72:665–78.
3. Yu-Feng Z, Yong H, Chao-Zhe Z, Qing-Jiu C, Man-Qiu S, Meng L, Li-Xia T,
Tian-Zi J, Yu-Feng W. Altered baseline brain activity in children with ADHD
revealed by resting-state functional MRI. Brain Dev. 2007;29:83–91.
4. COBRE database. http://fcon_1000.projects.nitrc.org/indi/retro/cobre.
html.
P162
Emergence of narrowband LFP oscillations from completely
asynchronous activity during seizures and high‑frequency
oscillations
Stephen V. Gliske1, William C. Stacey1,2, Eugene Lim3, Katherine A.
Holman4, Christian G. Fink3,5
1
Department of Neurology, University of Michigan, Ann Arbor, MI 48104,
USA; 2Department of Biomedical Engineering, University of Michigan,
Ann Arbor, MI 48104, USA; 3Department of Physics, Ohio Wesleyan
University, Delaware, OH 43015, USA; 4Department of Physics, Towson
University, Towson, MD 21252, USA; 5Neuroscience Program, Ohio
Wesleyan University, Delaware, OH 43015, USA
Correspondence: Christian G. Fink ‑ cgfink@owu.edu
BMC Neuroscience 2016, 17(Suppl 1):P162
Recent experimental studies have demonstrated the emergence of
narrowband local field potential oscillations during epileptic seizures
in which the underlying neural activity appears to be completely asynchronous [1]. We derive a mathematical model explaining how this
counterintuitive phenomenon may occur, showing that a population
of independent, completely asynchronous neurons may produce narrowband oscillations if each neuron fires quasi-periodically. This quasiperiodicity can occur through cells with similar frequency–current (f–I)
curves receiving a similar, high amount of uncorrelated synaptic noise.
Thus, this source of oscillatory behavior is distinct from the usual cases
(pacemaker cells entraining a network, or oscillations being an inherent property of the network structure), as it requires no oscillatory
drive nor any specific network or cellular properties other than cells
that repetitively fire with continual stimulus.
We deduce bounds on the degree of variability in neural spike-timing
which will permit the emergence of such oscillations, both for action
potential- and postsynaptic potential-dominated LFPs. (See Fig. 86
for example voltage traces and energy spectra resulting from asynchronous neural activity, demonstrating how our model naturally
explains why PSPs tend to dominate the LFP at low frequency, while
APs dominate at high frequency.) These results suggest that even an
uncoupled network may generate collective rhythms, implying that
the breakdown of inhibition and high synaptic input often observed
BMC Neurosci 2016, 17(Suppl 1):54
Page 93 of 112
during epileptic seizures may generate narrowband oscillations. We
propose that this mechanism may explain why so many disparate epileptic pathologies can produce similar high frequency oscillations [2].
References
1. Truccolo W, et al. Neuronal ensemble synchrony during human focal
seizures. J Neurosci. 2014;34:9927–44.
2. Engel Jr J, Bragin A, Staba R, Modi I. High-frequency oscillations: what is
normal and what is not? Epilepsia. 2009;50:598–604.
P163
Neuronal diversity in structure and function: cross‑validation
of anatomical and physiological classification of retinal ganglion
cells in the mouse
Jinseop S. Kim1,2, Shang Mu2, Kevin L Briggman3, H. Sebastian Seung2,4,
and the EyeWirers5
1
Department of Structure and Function of Neural Networks, Korea
Brain Research Institute, Daegu 41068, Republic of Korea; 2Princeton
Neuroscience Institute, Princeton University, Princeton, NJ 08544,
USA; 3Circuit Dynamics and Connectivity Unit, National Institute
of Neurological Disorders and Stroke, National Institutes of Health,
Bethesda, MD 20824, USA; 4Computer Science Department, Princeton
University, Princeton, NJ 08544, USA; 5http://eyewire.org
Correspondence: Jinseop S. Kim ‑ jinseop.s.kim@kbri.re.kr
BMC Neuroscience 2016, 17(Suppl 1):P163
The neural computation of visual perception begins in the retina. The
retinal neural circuits receive inputs from the photoreceptors, spread
out along interneurons, and converge to retinal ganglion cells (RGCs).
The axons of RGCs are the only output of the retina and carry all the
visual information from the retina to the rest of the brain. Each type of
RGCs is thought to be associated with one microcircuit and to process
distinct visual information. Therefore, classifying the types is an important step towards understanding the neural computation in the retina
and retina’s role in vision [1, 2].
We anatomically classified roughly 400 RGCs based mainly on dendritic stratification profiles [3]. The RGC dendritic arbors were reconstructed from serial electron microscope (EM) images of a (0.3 mm)2
slice of the inner plexiform layer of the mouse retina [4]. The reconstruction was carried out on EyeWire, a web-based EM reconstruction
pipeline that combines artificial intelligence of deep learning and
human intelligence of a community of ‘citizen neuroscientists’ [5]. This
is the first time EM reconstruction was done on a large enough area to
potentially sample and identify all RGC types.
For cross-validation of the anatomical classification, we compared it
with the visual responses of the same cells recorded by calcium imaging performed before EM preparation. The comparison confirmed that
our classification recovered all well-known ganglion cell types including on–off direction selective ganglion cells (DSGCs), sustained/transient On DSGCs, asymmetric Off DSGC types, sustained/transient and
On/Off alpha cells, and local edge detectors. We also found orientation
selective or direction selective responses in some cell types that were
not previously well-characterized or were previously unknown. In all,
our classification includes over 40 types of RGCs.
References
1. Sanes JR, Masland RH. The types of retinal ganglion cells: current
status and implications for neuronal classification. Annu Rev Neurosci.
2015;38:221–46.
2. Baden T, Berens P, Franke K, Rosón MR, Bethge M, Euler T. The functional
diversity of retinal ganglion cells in the mouse. Nature. 2016;529:345–50.
3. Sümbül U, Song S, McCulloch K, Becker M, Lin B, Sanes JR, Masland RH,
Seung HS. A genetic and computational approach to structurally classify
neuronal types. Nat Commun. 2014;5:3512.
4. Briggman KL, Helmstaedter M, Denk W. Wiring specificity in the directionselectivity circuit of the retina. Nature. 2011;471:183–88.
5. Kim JS, Greene MJ, Zlateski A, Lee K, Richardson M, Turaga SC, Purcaro M,
Balkam M, Robinson A, Behabadi BF, et al. Space–time wiring specificity
supports direction selectivity in the retina. Nature. 2014;509:331–6.
Fig. 86 Normalized energy spectra and voltage traces resulting from
asynchronous neural activity. A, B Results of superimposed, asynchronous action potential waveforms for quasi-periodic frequencies of
100 Hz (A) and 200 Hz (B). C, D Results of superimposed, asynchronous postsynaptic potential waveforms for quasi-periodic frequencies of 100 Hz (A) and 200 Hz (B). Gray dashed lines represent energy
spectra that would result from Poisson process spike trains convolved
with AP/PSP waveforms
P164
Analysis and modelling of transient firing rate changes in area MT
in response to rapid stimulus feature changes
Detlef Wegener1, Lisa Bohnenkamp1,2, Udo A. Ernst2
1
Brain Research Institute, University of Bremen, 28334 Bremen, Germany;
2
Institute for Neurophysics, University of Bremen, 28334 Bremen,
Germany
Correspondence: Detlef Wegener ‑ wegener@brain.uni‑bremen.de
BMC Neuroscience 2016, 17(Suppl 1):P164
Neurons in area MT of the primate visual system are strongly tuned
to the direction and speed of moving stimuli, and they exhibit pronounced transients in their firing rates after changes in visual stimulation. These transients increase the sensitivity of neurons and they
are closely correlated to behavioral performance. For example, arbitrary instantaneous speed changes are associated with transients of
different sign and amplitude, which closely correlate with the sign
and magnitude of the preceding stimulus change and with behavioral performance [1, 2]. Interestingly, the transients’ size cannot be
directly referred from the neuron’s underlying speed tuning, and is
significantly more pronounced if the base speed before the change
is far from the neuron’s preferred speed. Understanding the neural
dynamics shaping these responses, and their effects on information
transmission of arbitrary time-varying signals, is key to understanding
how the visual system copes with dynamic scenes. We here present a
dynamical model for MT neurons that reproduces detailed characteristics of experimentally observed transients (Fig. 87). The model takes
the single cell’s kinetics and its speed tuning into account. Based on
BMC Neurosci 2016, 17(Suppl 1):54
Page 94 of 112
References
1. Galashan FO, Saßen HC, Kreiter AK, Wegener D. Monkey area MT latencies
to speed changes depend on attention and correlate with behavioral
reaction times. Neuron. 2013;78(4):740–50.
2. Traschütz A, Kreiter AK, Wegener D: Transient activity in monkey area
MT represents speed changes ind is correlated with human behavioral
performance. J Neurophysiol 2015, 113(3):890-903.
P165
Step‑wise model fitting accounting for high‑resolution spatial
measurements: construction of a layer V pyramidal cell model
with reduced morphology
Tuomo Mäki‑Marttunen1, Geir Halnes2, Anna Devor3,4, Christoph
Metzner5, Anders M. Dale3,4, Ole A. Andreassen1, Gaute T. Einevoll2,6
1
NORMENT, Institute of Clinical Medicine, University of Oslo, Norway;
2
Department of Mathematical Sciences and Technology, Norwegian
University of Life Sciences, Ås, Norway; 3Department of Neurosciences,
University of California San Diego, La Jolla, CA, USA; 4Department
of Radiology, University of California San Diego, La Jolla, CA, USA;
5
Biocomputation Research Group, University of Hertfordshire, Hatfield,
UK; 6Department of Physics, University of Oslo, Norway
Correspondence: Tuomo Mäki‑Marttunen ‑ tuomomm@uio.no
BMC Neuroscience 2016, 17(Suppl 1):P165
Fig. 87 A, B Fits to motion onset responses to estimate each
neuron’s kinetics. C Experimentally estimated MT transients to positive and negative speed changes of various magnitude. D Transient
response amplitudes as derived from the model. E, F Relation
between transient and sustained MT responses as a function of speed
change magnitude as estimated experimentally (E) and by the model
(F)
divisive inhibition of excitation, it is capable to reproduce and explain
the specific transients of single neurons. Single direction column are
made up of one excitatory and one inhibitory population, with the
inhibitory population providing divisive inhibition onto the excitatory
population. By combining multiple direction columns to one hypercolumn, the model consists of N × 2 populations, with the excitatory populations receiving different input depending on their tuning
parameters and the stimulus, and the inhibitory populations receiving
an input averaged over the neighboring columns’ input by a Gaussian
kernel plus a fixed offset. Using an optimization procedure, the model
reliably reproduces MT cell responses to arbitrary accelerations and
decelerations of a moving stimulus, starting from both low and high
base speeds, reproducing recently unexplained experimental data. If
the inhibitory time constant is a multiple of the excitatory time constant, the model is analytically tractable for a piecewise constant input
current: The analytical solution allows quantifying the transients’ magnitude as a function of general neuron parameters such as response
gain and time constants, providing precise predictions for population
responses to brief events of arbitrary contrast.
Novel imaging methods such as intracellular Ca2+ imaging and voltage-sensitive dye measurements provide ever finer spatiotemporal
data about single-neuron activity. The challenge for model fitting
methods is to incorporate these data in order to describe the neuron
behavior in a manner that faithfully preserves the signal propagation
and membrane potential dynamics across the neuronal dendrites.
A difficulty in this task is the evidently large number of different ion
channels residing along the dendritic and perisomatic locations:
Unless extra care is taken, the role of specific species of ion channel
could be under- or overestimated at the expense of another type of
ion channel.
In this work, we propose an automatic step-wise model fitting procedure as a solution to this challenge. Our approach resembles that of
[1], but our objective functions are designed to account for correct
membrane potentials not only at soma but also along the dendrites. In
addition, we replace the need for spatial occlusion of parts of dendrite
(“pinching”) [2] in the experimental setup by a cumulative use of ion
channel blockers.
We apply this procedure to construct a reduced-morphology version
of the layer V pyramidal cell model of [3]. We simulated the cumulative blocking of ion channels by setting the corresponding ion channel
conductances to zero in the full model, and measured the membrane
potential (and Ca2+ concentrations when needed) along the soma and
dendrites at each step. We then fitted the maximal conductances in
the model with reduced morphology in four steps, starting with passive parameters (1st step), continuing with Ih current conductances
(2nd step), Ca2+ dynamics and related conductances (3rd step), and
ending with ion channel conductances that are in charge of the spiking behavior (4th step).
We show that our model with reduced morphology correctly reproduces important aspects of the membrane potential dynamics across
the neuron, both in the control condition (see Fig. 88), and under the
effect of the abovementioned ion channel blockers. In the final step
of our study, we present and apply a method for reducing the number of synaptic contacts (from 1000s to a few 100s) yet maintaining
the spatio-temporal activation pattern of the neuron. The obtained
network model is cost-efficient in terms of both simulation time and
memory requirements. Our model is publicly accessible in ModelDB,
accession number 187474, as NEURON and NeuroML-2 descriptions
(https://senselab.med.yale.edu/ModelDB/showModel.cshtml?mo
del=187474).
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 88 Comparison of model with reduced (red) morphology to the
model with full (blue) morphology. The y-axis shows the membrane
potential at soma (solid) and apical dendrite (dashed) as a response to
a somatic 200-ms DC pulse
References
1. Bahl A, Stemmler MB, Herz AV, Roth A. Automated optimization of a
reduced layer 5 pyramidal cell model based on experimental data. J
Neurosci Methods. 2012;210(1):22–34.
2. Bekkers JM, Häusser M. Targeted dendrotomy reveals active and passive
contributions of the dendritic tree to synaptic integration and neuronal
output. Proc Natl Acad Sci. 2007;104(27):11447–52.
3. Hay E, Hill S, Schürmann F, Markram H, Segev I. Models of neocortical
layer 5b pyramidal cells capturing a wide range of dendritic and perisomatic active properties. PLoS Comput Biol. 2011;7:e1002107.
P166
Contributions of schizophrenia‑associated genes to neuron firing
and cardiac pacemaking: a polygenic modeling approach
Tuomo Mäki‑Marttunen1, Glenn T. Lines2, Andy Edwards2, Aslak Tveito2,
Anders M. Dale3, Gaute T. Einevoll4, Ole A. Andreassen1
1
NORMENT, Institute of Clinical Medicine, University of Oslo, Norway;
2
Simula Research Laboratory and Center for Cardiological Innovation,
Oslo, Norway; 3Multimodal Imaging Laboratory, UC San Diego, La Jolla,
CA, USA; 4Department of Mathematical Sciences and Technology,
Norwegian University of Life Sciences, Ås, Norway
Correspondence: Tuomo Mäki‑Marttunen ‑ tuomomm@uio.no
BMC Neuroscience 2016, 17(Suppl 1):P166
A recent genome-wide association study (GWAS) of schizophrenia
(SCZ) has identified more than a hundred genetic loci exceeding
genome-wide significance, confirming the polygenic nature of the
disorder [1]. The loci implicate genes that encode numerous ion channel subtypes and calcium transporters, and are major contributors
not only to the function of brain cells, but also to the functioning of
organs outside the central nervous system, such as heart. Meta-studies have reported a 2.5-fold–threefold increase in mortality rates in
schizophrenic patients, and majority of these excess deaths are natural
and mostly due to cardiovascular disease [2]. In agreement with this
observation, GWASs of cardiac phenotypes, such as electrocardiographic (ECG) measures, highlight a set of genes that overlaps with the
one discovered in GWASs of SCZ. Nevertheless, both the genetic and
mechanistic connections between cardiac and neural phenotypes in
SCZ patients remain poorly understood.
In this work, we use computational modeling to study the contribution of SCZ-associated genes to cardiac and neuronal excitability. We
focus our analyses on two central, well-studied cell types, namely,
layer V pyramidal cells (L5PCs) in the cortex and sinoatrial node cells
Page 95 of 112
(SANCs) in the myocardium. The apical tuft of an L5PC serves as an
integration hub for non-local synaptic inputs, and is considered a biological substrate for cortical associations providing high-level “context” for low-level (e.g., sensory) inputs that arrive to the perisomatic
compartment. Therefore, the ability of L5PC to integrate the apical
and perisomatic inputs has been proposed as one of the mechanisms
that could be impaired in hallucinating patients. The SANCs, in turn,
have a key role in controlling the heart rate as the primary pacemakers
of the mammalian heart. Both of these cell types are well described
in terms of biophysical modeling, and are therefore a suitable target
for a detailed computational studies incorporating genetic effects. We
apply two recent multicompartmental L5PC models and two recent
SANC models to argue for the generality of our findings.
We show that small changes in the parameters governing the voltagedependence and time constants of activation and inactivation of different ion channels caused observable effects in both L5PC and SANC
function. In the case of Ca2+ channel gene variants, these changes
typically had opposite effects on cell excitability in L5PCs compared
to SANCs (higher L5PC firing frequency ↔ lower SANC pacemaking frequency), while in the case of Na+ or HCN channel variants, the
effects were mostly similar (higher L5PC firing frequency ↔ higher
SANC pacemaking frequency). Furthermore, many of the studied variants showed an impact on signal propagation in a chain of coupled
SANCs. Our results may help explain some of the cardiac comorbidity
in schizophrenia, and may facilitate generation of effective antipsychotic medications with less arrythmia side-effects.
References
1. Ripke S, Sanders AR, Kendler KS, Levinson DF, Sklar P, Holmans PA, Lin DY,
Duan J, Ophoff RA, Andreassen OA et al.: Genome-wide association study
identifies five new schizophrenia loci. Nat Gen. 2011;43:969–76.
2. Laursen TM, Munk-Olsen T, Vestergaard M. Life expectancy and cardiovascular mortality in persons with schizophrenia. Curr Opin Psychiatry.
2012;25(2):83–8.
3. Hay E, Hill S, Schürmann F, Markram H, Segev I: models of neocortical
layer 5b pyramidal cells capturing a wide range of dendritic and perisomatic active properties. PLoS Comput Biol. 2011;7:e1002107.
4. Almog M, Korngreen A. A quantitative description of dendritic conductances and its application to dendritic excitation in layer 5 pyramidal
neurons. J Neurosci. 2014;34(1):182–96.
5. Kharche S, Yu J, Lei M, Zhang H. A mathematical model of action potentials of mouse sinoatrial node cells with molecular bases. Am J Physiol
Heart Circ Physiol. 2011;301(3):H945–63.
6. Severi S, Fantini M, Charawi LA, DiFrancesco D. An updated computational model of rabbit sinoatrial action potential to investigate the
mechanisms of heart rate modulation. J Physiol. 2012;590(18):4483–99.
P167
Local field potentials in a 4 × 4 mm2 multi‑layered network model
Espen Hagen1, Johanna Senk1, Sacha J van Albada1, Markus Diesmann1,2,3
1
Institute of Neuroscience and Medicine (INM‑6) and Institute
for Advanced Simulation (IAS‑6) and JARA BRAIN Institute I, Jülich
Research Centre, Jülich, 52425, Germany; 2Department of Psychiatry,
Psychotherapy and Psychosomatics, Medical Faculty, RWTH Aachen
University, Aachen, 52074, Germany; 3Department of Physics, Faculty 1,
RWTH Aachen University, Aachen, 52074, Germany
Correspondence: Espen Hagen ‑ e.hagen@fz‑juelich.de
BMC Neuroscience 2016, 17(Suppl 1):P167
The local field potential (LFP), the low-frequency part of extracellular potentials in neural tissue, is routinely recorded as a measure of
population activity. LFPs reflect correlated activity of both local and
remote neurons and depend on the anatomy and electrophysiology
of neurons near the recording location. While forward models have
shed light on various aspects of LFPs, e.g., their spatial reach [1], such
models often ignore network interactions. Large-scale network models
commonly use point neurons for tractability (see, e.g., [2]). However,
predicting the LFP signal from such models is not straightforward, as
point neurons do not generate extracellular potentials. In [3] we provided methods to compute extracellular potentials from point-neuron
networks incorporating the biophysical principles of LFP generation
BMC Neurosci 2016, 17(Suppl 1):54
Page 96 of 112
4
McConnell Brain Imaging Centre, Montreal Neurological Institute,
McGill University, Montreal, Canada; 5Department of Computational
Neuroscience, University Medical Center Eppendorf, Hamburg, Germany;
6
Department of Health Sciences, Boston University, Boston, MA, USA;
7
Department of Psychiatry, Psychotherapy and Psychosomatics, Medical
Faculty, RWTH Aachen University, Aachen, Germany; 8Department
of Physics, Faculty 1, RWTH Aachen University, Aachen, Germany
Correspondence: Maximilian Schmidt ‑ max.schmidt@fz‑juelich.de
BMC Neuroscience 2016, 17(Suppl 1):P168
Fig. 89 A Instantaneous spiking and LFP in a 4-layer network
model covering 4 × 4 mm2 at realistic cell and synapse density with
distance-dependent connectivity. B Pairwise correlations between
spike trains of exc. (E) and inh. (I) layer 5 neurons as function of
distance (red: E–E, blue: I–I, black: E–I). C Distance-dependent LFP
correlation computed for a 10 × 10 electrode grid in layer 5 (0.4 mm
between contacts)
using multicompartment neurons. This hybrid scheme uses spike times
of point neurons as spatially dependent synaptic input with layer specificity of connections from anatomical data. The methods were demonstrated using a laterally homogeneous, layered point-neuron network
representing 1 mm2 of early sensory cortex at full cell and synapse
density [4]. Preserving biological cell and connection densities is critical: networks may not be strongly downscaled without affecting correlations [5], and diluted LFP-generating populations fail to preserve
the effect of correlations on the LFP [3]. Even small network correlations dominate in the compound LFP spectrum due to the different
scaling of average single-cell LFP spectra and average pairwise coherence of single-cell LFP. Here, we extend this work to a network covering 4 × 4 mm2 (Fig. 89A) accounting for connection probabilities falling
off with lateral distance. Even for low pairwise spike-train correlations
(Fig. 89B), the model accounts for highly correlated LFPs across lateral
distance (Fig. 89C) as observed experimentally. Further we show that
such features strongly depend on network state.
Acknowledgements: EU FP7 grant 604102 (HBP); Helmholtz Portfolio
Supercomputing and Modeling for the Human Brain (SMHB).
References
1. Lindén H, Tetzlaff T, Potjans TC, Pettersen KH, Gruen S, Diesmann M, Einevoll GT. Modeling the spatial each of the LFP. Neuron. 2011;72:859–72.
2. Bos H, Diesmann M, Helias M. Identifying anatomical origins of coexisting
oscillations in the cortical microcircuit. arXiv:1510.00642 [q-bio.NC] 2016.
3. Hagen E, Dahmen D, Stavrinou ML, Lindén H, Tetzlaff T, van Albada SJ,
Grün S, Diesmann M, Einevoll GT. Hybrid scheme for modeling local field
potentials from point-neuron networks. arXiv:1511.01681 [q-bio.NC]
2016. http://inm-6.github.io/hybridLFPy.
4. Potjans TC, Diesmann M. The cell-type specific cortical microcircuit:
relating structure and activity in a full-scale spiking network model. Cereb
Cortex. 2014;24:785–806.
5. Van Albada SJ, Helias M, Diesmann M. Scalability of asynchronous networks is limited by one-to-one mapping between effective connectivity
and correlations. PLoS Comput Biol. 2015;11(9):e1004490.
P168
A spiking network model explains multi‑scale properties
of cortical dynamics
Maximilian Schmidt1, Rembrandt Bakker1,2, Kelly Shen3, Gleb Bezgin4,
Claus‑Christian Hilgetag5,6, Markus Diesmann1,7,8, Sacha Jennifer van
Albada1
1
Institute of Neuroscience and Medicine (INM‑6) and Institute
for Advanced Simulation (IAS‑6), and JARA BRAIN Institute I, Jülich
Research Centre, Jülich, Germany; 2Donders Institute for Brain, Cognition
and Behavior, Radboud University, Nijmegen, Netherlands; 3Rotman
Research Institute, Baycrest, Toronto, Ontario M6A 2E1, Canada;
Neural networks in visual cortex are structured into areas, layers, and
neuronal populations with specific connectivity at each level. Cortical dynamics can similarly be characterized on different scales, from
single-cell spiking statistics to the structured patterns of interactions
between areas. A challenge of computational neuroscience is to investigate the relation of the structure of cortex to its dynamics. Network
models are promising tools, but for technical and methodological reasons, they have been restricted to detailed models of one or two areas
or large-scale models that reduce the internal structure of areas to a
small number of differential equations.
We here present a multi-scale spiking network model of all visionrelated areas of macaque cortex that represents each area by a fullscale microcircuit with area-specific architecture based on a model
of early sensory cortex [1]. The layer- and population-resolved network connectivity integrates axonal tracing data from the CoCoMac
database with recent quantitative tracing data, and is systematically refined using dynamical constraints [2]. Gaps in the data are
bridged by exploiting regularities of cortical structure such as the
exponential decay of connection densities with inter-areal distance
and a fit of laminar patterns versus logarithmized ratios of neuron
densities.
Simulations reveal a stable asynchronous irregular ground state with
heterogeneous activity across areas, layers, and populations. In the
presence of large-scale interactions, the model reproduces longer
intrinsic time scales in higher compared to early visual areas, similar to
experimental findings [3]. Activity propagates preferentially in the feedback direction, mimicking experimental results associated with visual
imagery [4]. Cortico-cortical interaction patterns agree well with fMRI
resting-state functional connectivity [5]. The model bridges the gap
between local and large-scale accounts of cortex, and clarifies how the
detailed connectivity of cortex shapes its dynamics on multiple scales.
Acknowledgements: VSR computation time Grant JINB33, Helmholtz Portfolio SMHB, EU Grant 269921 (BrainScaleS), EU Grant 604102
(Human Brain Project, HBP), SFB936/A1, Z1 and TRR 169/A2.
References
1. Potjans TC, Diesmann M. The cell-type specific cortical microcircuit:
relating structure and activity in a full-scale spiking network model. Cereb
Cortex. 2014;24:785–806.
2. Schuecker J, Schmidt M, van Albada SJ, Diesmann M, Helias M. Fundamental activity constraints lead to specific interpretations of the connectome. arXiv preprint 2015, arXiv:1509.03162.
3. Murray JD, Bernacchia A, Freedman DJ, Romo R, Wallis JD, Cai X, PadoaSchioppa C, Pasternak T, Seo H, Lee D, Wang X-J. A hierarchy of intrinsic
timescales across primate cortex. Nat Neurosci. 2014;17:1661–3.
4. Dentico D, Cheung BL, Chang J-Y, Guokas J, Boly M, Tononi G, Van Veen B.
Reversal of cortical information flow during visual imagery as compared
to visual perception. Neuroimage. 2014;100:237–243.
5. Shen K, Bezgin G, Hutchison RM, Gati JS, Menon RS, Everling S, McIntosh
AR. Information processing architecture of functionally defined clusters in
the macaque cortex. J Neurosci. 2012;32:17465–76.
P169
Using joint weight‑delay spike‑timing dependent plasticity
to find polychronous neuronal groups
Haoqi Sun1,2,3,5, Olga Sourina2,5, Guang‑Bin Huang3,5, Felix Klanner4,5,
Cornelia Denk5
1
Energy Research Institute @ NTU (ERI@N), Interdisciplinary Graduate
School, Nanyang Technological University, Singapore 639798; 2Fraunhofer
IDM @ NTU, Nanyang Technological University, Singapore 639798;
BMC Neurosci 2016, 17(Suppl 1):54
3
School of Electrical and Electronic Engineering, Nanyang Technological
University, Singapore 639798; 4School of Computer Engineering,
Nanyang Technological University, Singapore 639798; 5Future Mobility
Research Lab, A Joint Initiative of BMW Group & NTU, Nanyang
Technological University, Singapore 639798
Correspondence: Haoqi Sun ‑ hsun004@e.ntu.edu.sg
BMC Neuroscience 2016, 17(Suppl 1):P169
It is known that polychronous neuronal groups (PNGs), i.e. neuron
groups having reproducible time-locked but not synchronous firing
patterns, can function as representative entities [1]. They have huge
capacity by sharing neurons. They compete between each other to
represent sensory inputs. Therefore, PNG is considered as one of the
potential, yet elusively difficult to analyze, hypothetical mechanisms of
memory in the brain.
In computational models, the difficulties of finding PNGs mainly come
from (1) low percentage of spikes from PNGs (about 4 % [1]) when
driven by random inputs; and (2) combination explosion to enumerate
all possible PNGs for template-matching (possible PNGs triggered by 3
neurons in a 1000-neuron network is 3C1000 = 1.66 × 108).
Here we aim at solving the second difficulty without template-matching by connecting PNG readout neurons with joint weight-delay spiketiming dependent plasticity (joint STDP) to the network. The joint
STDP consists of (1) weight STDP with the conventional exponential
learning window; and (2) (axonal) delay STDP with learning window of
shape te−t/τ, scaled by weight-related gains. The joint STDP strengthens and pulls together spikes arriving before postsynaptic firing, on
the other hand weakens and postpones spikes after postsynaptic firing. In this way, we can recover the PNG by looking at (1) the strengthened synapses, which tells which neurons belong to the PNG; and (2)
the delays of the strengthened synapses, which are complementary to
the spike timing inside the PNG, because the presynaptic spike arrival
times for the readout neuron (=spike timing + delay) are pulled close
to each other.
In the experiment, we repeatedly fed structured inputs to a sparsely
connected network of 800 excitatory and 200 inhibitory neurons.
There were 150 readout neurons connected to the network with lateral inhibition between them. After 405 s of simulation, we used the
incoming weights and delays of the readout neurons to find PNGs (see
Fig. 90). It turned out that the readout neurons can learn the subsets
of the persistently activated PNGs. The readout neurons do not rely on
template-matching. Instead, they become differentiated members of
the PNG to indicate the actual activation of its subsets.
Page 97 of 112
Reference
1. Izhikevich EM. Polychronization: computation with spikes. Neural Comput. 2006;18(2):245–82.
P170
Tensor decomposition reveals RSNs in simulated resting state
fMRI
Katharina Glomb1, Adrián Ponce‑Alvarez1, Matthieu Gilson1, Petra
Ritter2,3,4,5, Gustavo Deco1,6
1
Center for Brain and Cognition, Universitat Pompeu Fabra, 08018
Barcelona, Spain; 2Minerva Research Group Brain Modes, Max Planck
Institute for Human Cognitive and Brain Sciences, 04103 Leipzig,
Germany; 3Department of Neurology, Charité ‑ University Medicine,
10117 Berlin, Germany; 4Bernstein Focus State Dependencies of Learning
& Bernstein Center for Computational Neuroscience, 10115 Berlin,
Germany; 5Berlin School of Mind and Brain & Mind and Brain Institute,
Humboldt University, 10117 Berlin, Germany; 6Catalan Institution
for Advanced Studies (ICREA), Universitat Barcelona, 08010 Barcelona,
Spain
Correspondence: Won Hee Lee ‑ katharina.glomb@upf.edu
BMC Neuroscience 2016, 17(Suppl 1):P170
The subject of this study is the temporal dynamics of functional connectivity (FC) in human resting state (RS) as measured with BOLD fMRI.
In spite of rising interest in the topic [1], it remains unclear whether
observed FC is stationary or if state switching is present, nor is it clear
what constitutes these putative states. Modelling is an invaluable tool
for answering these questions: here we combine a dynamic mean field
model of the cortex with data analysis in order to determine whether
and to what extent spatio-temporal FC patterns found in empirical data
can be mimicked by a stationary model as described in [2]. To this end,
we cast our data into tensor form by computing time-dependent FC
inside of sliding windows (dynamic FC, dFC), comparing three methods
to compute dFC (two correlation based, and mutual information). We
employ canonical polyadic decomposition (also known as parallel factor analysis) with or without non-negativity constraint to decompose
the tensors, which allows us to simultaneously consider the temporal
and spatial dimensions [3]. First, we decompose such tensors obtained
from empirical data of 24 subjects [4] and cluster resulting spatial features (i.e., communities) in order to obtain a small number of templates.
These templates are used in a second step to compare to simulated
data that is processed in the same way. We find that even on a very
low level of spatial resolution (66 cortical regions), and using only the
2 % biggest dFC values in terms of region pairs and time windows, we
succeed in extracting communities that generalize across subjects and
can be found in the simulated data. Furthermore, we show that using
model-based effective connectivity to inform the model [5] leads to
more realistic and stable communities than diffusion weighted MRIbased structural connectivity alone. The method shown here is widely
applicable to compare patient groups, data obtained from different
tasks as well as mental states, and opens the door to understanding the
differences between the temporal dynamics of these conditions.
Acknowledgements: KG is funded by the Marie Curie Initial Training
Network INDIREA, grant agreement no ITN-2013-606901. APA was
supported by SEMAINE ERA-Net NEURON Project. MG acknowledges
funding from FP7 FET ICT Flagship Human Brain Project (604102). GD
was supported by the ERC Advanced Human Brain Project (n. 604102)
and the Plan Estatal de Fomento de la investigación Científica y Técnica de Excelencia (PSI2013-42091-P).
Fig. 90 A The spike raster plot showing 0.6 s of simulation. The
vertical axis shows neuron index. Neurons from index 1 to 100 receive
structured inputs. Colored spikes refer to PNGs founded by the same
colored spikes of readout neurons (above the dash line), where the
letter-marked ones are shown in other panels. B A recovered PNG
with the predicted spike timing (receptive field) and the actual spikes.
The numbers are neuron indices. C The same PNG in B but activated
at another time. D, E Another PNG
References
1. Hutchison RM, Womelsdorf T, Allen EA, Bandettini PA, Calhoun VD,
Corbetta M, Della Penna S, Duyn JH, Glover GH, Gonzalez-Castillo J,
Handwerker DA, Keilholz S, Kiviniemi V, Leopold DA, de Pasquale F, Sporns
O, Walter M, Chang C. Dynamic functional connectivity: promise, issues,
and interpretations. NeuroImage. 2013;80:360–78.
2. Deco G, Ponce-Alvarez A, Hagmann P, Romani G, Mantini D, Corbetta M.
How local excitation–inhibition ratio impacts the whole brain dynamics. J
Neurosci. 2014;34(23):7886–98.
BMC Neurosci 2016, 17(Suppl 1):54
3.
4.
5.
Cichocki A. Tensor decompositions: a new concept in brain data analysis?
arxiv Prepr. 2013, arXiv1305.0395, 507–17.
Schirner M, Rothmeier S, Jirsa VK, McIntosh AR, Ritter P. An automated
pipeline for constructing personalized virtual brains from multimodal
neuroimaging data. NeuroImage. 2015;117:343–57.
Gilson M, Moreno-Bote R, Ponce-Alvarez A, Ritter P, Deco G. Estimation of
directed effective connectivity from fMRI functional connectivity hints at
asymmetries in cortical connectome. PLoS Comput Biol. 2016.
P171
Getting in the groove: testing a new model‑based method
for comparing task‑evoked versus resting‑state activity in fMRI
data on music listening
Matthieu Gilson1,†, Maria A. G. Witek2,†, Eric F. Clarke3, Mads Hansen4,
Mikkel Wallentin5, Gustavo Deco1, Morten L Kringelbach2,5,6, Peter Vuust2,5
1
Center for Brain Cognition, Universitat Pompeu Fabra, Barcelona, Spain;
2
Center for Music in the Brain, Aarhus University & Royal Academy
of Music, Aarhus/Aalborg, Denmark; 3Faculty of Music, University
of Oxford, UK; 4Department of Psychology and Behavioural Sciences,
Aarhus University, Denmark; 5Center of Functionally Integrative
Neuroscience, Aarhus University, Denmark; 6Department of Psychiatry,
University of Oxford, UK
Correspondence: Matthieu Gilson ‑ matthieu.gilson@upf.edu
BMC Neuroscience 2016, 17(Suppl 1):P171
†
Equal contribution.
Much of present neuroimaging studies have used fMRI to simply
measure the activity in brain regions and computing the functional
connectivity (FC) between regions and behaviour. This has provided
important insights into the task-evoked activity compared to rest and
thus the flow of information between functional regions (e.g., sensory, multimodal integration, memory). Yet, this does not capture all
of the complex spatiotemporal patterns of brain activity and in particular not been the underlying effective connectivity. The present
study provides evidence for application of a novel method of determining the functional roadmap of effective connectivity (EC), which
measures the strengths of dynamic cortical interactions. We use a
whole-brain dynamical model that combines fMRI data with anatomical information obtained using diffusion-tensor imaging (DTI) [1]. Our
recently developed method [2] provides estimates for the EC as well
as the local excitability and stimulus load in a study of groove-based
music. The brain is divided into 90 areas and the input noise is shaped
by the EC to generate the FC. This model allows us to explore the
role of the network parameters in shaping FC: after constraining the
model to reproduce resting-state activity, we examine the effect of
an arbitrary change in individual inputs and EC strengths on FC. Our
method focuses on spatio-temporal FC, meaning covariances of BOLD
signals with possible time shifts. The estimated EC and inputs are
taken as fingerprints of the brain dynamics. We analyze fMRI data of
participants listening to 15 rhythms with three levels of syncopation:
Low, Medium and High (five drum-breaks in each level) [3]. In accordance with other studies, we find behaviourally that the Medium
level—with more “groove”—elicits the most pleasure and wanting to
move. We tune the model to reproduce the FC recorded for each syncopation level, as well as rest. We analyze significant changes for each
groove condition as compared to rest. FC for Medium syncopation
exhibits a faster shuffling between successive brain patterns of activity, linked to more metastability and corresponding maybe to the
subjective experience of higher pleasure. In addition, our model gives
a detailed functional neuroanatomy of dynamical changes in the
brain networks. Interestingly, Medium syncopation induces changes
in excitability in the basal ganglia, such as the pallidum and the caudate nucleus, which may be related to the increased desire for moving. Interestingly, significant changes are also observed in regions of
the orbitofrontal and anterior cingulate cortices, which have been
strongly implicated in the pleasure network [4]. Overall, our new
method has for the first time allowed us to uncover the network and
the corresponding effective connectivity of a highly pleasurable state
of groove; possibly even revealing the brain topography of eudaimonia, the sense of well-being.
Page 98 of 112
Acknowledgements: MG and GD were supported by the FP7 FET ICT
Flagship Human Brain Project (604102). GD was supported by the ERC
Advanced Grant: DYSTRUCTURE (n. 295129), by the Spanish Research
Project SAF2010-16085. MLK was supported by the ERC Consolidator Grant: CAREGIVING (n. 615539). MLK, MW and PV were supported
by the Center for Music in the Brain, funded by the Danish National
Research Foundation (DNRF117).
References
1. Deco G, Jirsa V, McIntosh A. Emerging concepts for the dynamical
organization of resting-state activity in the brain. Nat Rev Neurosci.
2011;12:43–56.
2. Gilson M, Moreno-Bote R, Ponce-Alvarez A, Ritter P, Deco G. Estimation of
directed effective connectivity from fmri functional connectivity hints at
asymmetries in cortical connectome. PLoS Comput Biol. 2016 (in press).
3. Witek MAG, Clarke E, Wallentin M, Kringelbach ML, Vuust P. Syncopation, body-movement and pleasure in Groove music. PLoS One.
2014;9:e94446.
4. Berridge KC, Kringelbach ML. Pleasure systems in the brain. Neuron.
2015;86:646–64.
P172
Stochastic engine for pathway simulation (STEPS) on massively
parallel processors
Guido Klingbeil1, Erik De Schutter1
1
Computational Neuroscience Unit, Okinawa Institute of Science
and Technology, 1919‑1 Tancha, Onna‑son, Kunigami‑gun, Okinawa
904‑0495, Japan
Correspondence: Guido Klingbeil ‑ guido.klingbeil@oist.jp
BMC Neuroscience 2016, 17(Suppl 1):P172
STEPS is a stochastic reaction–diffusion simulator. Its emphasis is on
accurately simulating signaling pathways in realistic morphologies [1].
It is becoming apparent that larger computational models are
demanded to either capture more such morphologies or to simulate
more complex systems. As an example, the dendrite calcium burst
model presented by Anwar et al. [2] requires approximately 285,000
sub-volumes with 15 diffusing molecular species and 20 reactions
per sub-volume. It required several weeks to compute a simulation of
500 ms.
Thus it is desirable to reduce the computational burden. Accelerators
such as graphics processing units (GPU) offer unprecedented computing performance and are now common amongst the fastest super
computers [3]. This project enables STEPS to benefit from the computational power of GPUs.
GPUs are massively parallel co-processors aggregating thousands
of simplified processing cores onto a single chip. They share many
characteristics with vector computers and a key challenge is that the
processing cores are not independent. Similar to vector computers
an operation is applied to a group of data elements rather than to the
individual data element. Furthermore, the programmer has to mitigate the memory hierarchy of GPUs. While memory with a high access
latency is, in general, abundant, fast memory space shared between
threads is small which may limit the size of the reaction system one is
able to simulate.
Previous research has shown that we can exploit the computational
power of GPUs to accelerate spatially homogenous stochastic simulations by two orders of magnitude while avoiding the limitation
imposed to the size of the reaction system to be simulated by the
small fast memory space [4].
STEPS implements a spatial version of Gillespie’s stochastic simulation
algorithm (SSA) computing reaction–diffusion systems on a mesh of
tetrahedral sub-volumes [1, 5]. Currently a parallelised multi-processor
version of STEPS is under development. Operator splitting techniques
allow to separate the reaction of molecules within a sub-volume from
the diffusion of molecules between them. This prevents computationally costly rollbacks in case of molecules diffusing between sub-volumes handled by different processors.
We develop a layered hybrid software architecture using both,
the classic central processing unit as well as GPUs, integrated into
STEPS applying GPU acceleration at the sub-volume level and
BMC Neurosci 2016, 17(Suppl 1):54
integrating them into a coherent spatial simulation using operator
splitting.
Our architecture will be a plug-in solution to STEPS not requiring any
changes to the interfaces towards the user or other software systems
of STEPS itself.
References
1. Hepburn I, et al.: STEPS: efficient simulation of stochastic reaction–diffusion models in realistic morphologies. BMC Syst Biol. 2012;6:36.
2. Anwar H, et al.: Stochastic calcium mechanisms cause dendritic calcium
spike variability. J Neurosci. 2013;33(40):15848–67.
3. TOP500 Supercomputer Site [http://www.top500.org].
4. Klingbeil, et al. Stochastic simulation of chemical reactions with cooperating threads on GPUs (in preparation).
5. Gillespie DT. Exact stochastic simulation of coupled chemical reactions. J
Phys Chem. 1977;81(25):2340–61.
P173
Toolkit support for complex parallel spatial stochastic reaction–
diffusion simulation in STEPS
Weiliang Chen1, Erik De Schutter1
Computational Neuroscience Unit, Okinawa Institute of Science
and Technology, Okinawa 904‑0411, Japan
Correspondence: Weiliang Chen ‑ w.chen@oist.jp
BMC Neuroscience 2016, 17(Suppl 1):P173
The studies of large neuronal pathway models with complex morphologies, such as our previous work on the stochastic effects of calcium
dynamics in Purkinje cells [1], present a great challenge to currently
available spatial stochastic reaction–diffusion simulators, for example,
STEPS [2], as the model scales and complexities quickly surpass the
capability any serial simulator can achieve.
One possible solution for this challenge is parallelization. At CNS2015
we reported a parallel implementation of STEPS which demonstrated
great speedup when simulating a reduced calcium burst model with a
tetrahedral cylinder [3], but it is clear that to explore the full potential
of our implementation, a larger scale simulation with more complex
geometry is required.
Here we extend our work by simulating the reduced calcium burst
model with a reconstructed Purkinje dendrite tree branch mesh.
Comparing to previous simulations with regular cylinder meshes, the
simulation with dendrite tree mesh requires several new support routines from the simulator. First of all, the simple axis based partitioning approach used in the cylinder simulations is no longer a good
partitioning solution due to the complex tree structure of the mesh.
A sophisticated mesh partitioning and validation solution is therefore necessary for the new simulation. Second, the new simulation
demands good regional annotation and data collection support as
calcium concentration, the main focus of the simulation, varies both
spatially and temporally. The parallel environment further increases
the difficulty of such support as the simulation is distributed over a
massive number of processors, and each annotated region may not
be completely simulated within a single processor. Furthermore, it is
necessary to minimize the user interface difference between serial and
parallel STEPS solvers and extend the STEPS visualization toolkit to
facilitate comparison with results from previous work.
In this poster we demonstrate the general procedure of converting a
serial STEPS simulation to its parallel counterpart, using the reduced
calcium burst model with complex tree mesh as example, and showcase new supporting toolkits developed for the procedure. We believe
that the presented procedure and toolkits will be helpful to STEPS
users in their future research.
References
1. Anwar H, Hepburn I, Nedelescu H, Chen W, De Schutter E. Stochastic
calcium mechanisms cause dendritic calcium spike variability. J Neurosci.
2013;33(40):15488–867.
2. Hepburn I, Chen W, Wils S, De Schutter E. STEPS: efficient simulation of
stochastic reaction–diffusion models in realistic geometries. BMC Syst
Biol. 2012;6:36.
Page 99 of 112
3.
Chen W, Hepburn I, De Schutter, E. Implementation of parallel spatial
stochastic reaction–diffusion simulation in STEPS. BMC Neurosci.
2015;16(Suppl 1):P54.
P174
Modeling the generation and propagation of Purkinje cell
dendritic spikes caused by parallel fiber synaptic input
Yunliang Zang1, Erik De Schutter1
Computational Neuroscience Unit, Okinawa Institute of Science
and Technology Graduate University, Onna‑son, Okinawa, Japan
Correspondence: Yunliang Zang ‑ yunliang.zang@oist.jp
BMC Neuroscience 2016, 17(Suppl 1):P174
The dendrite of Purkinje cell (PC) has been shown to express different types of voltage gated ion channels. After strong parallel fiber (PF)
stimulus, calcium currents can cause dendritic spikes to occur in the
spiny dendrite [1]. Different with climbing fiber caused calcium signals that propagate throughout the dendritic tree, PF caused dendritic
spikes are local. The elevated calcium concentration due to the local
dendritic spike may trigger local synaptic plasticity, possibly playing
a significant role in information processing. However, until now, how
these dendritic spikes originate and propagate is not well understood.
In this work, we built a new PC dendrite model, which can generate
local dendritic calcium spikes. The generated spike by model shows
similar properties to experimental observations [1], including spike
threshold, amplitude and latency. We identify the role of P type Ca2+
current, A type K+ current, high threshold K+ current (Kv3), calcium
activated K+ current and axial current on the depolarization and repolarization of the spike. In the model, the required threshold synaptic
input to trigger local dendritic spikes decreases with distance from
soma, which facilitates the occurrence of spikes in the spiny dendrite
by PF synaptic input. This model can also successfully replicate the failure of propagation of PF caused dendritic spikes at the parent branch
point. By analyzing the spatial spread of the dendritic spikes and EPSP
signal to soma, we identify the relative contribution of active currents
and impedance mismatch on the signal decay. Because dendritic
spikes can robustly propagate over an entire branchlet in the direction
away from the soma, dendritic branchlets may be the basic organization unit for integrating synaptic input [2, 3].
References
1. Rancz EA, Hausser M: Dendritic calcium spikes are tunable triggers of
cannabinoid release and short-term synaptic plasticity in cerebellar
Purkinje neurons. J Neurosci. 2006;26(20):5428–37.
2. De Schutter E, Bower JM. Simulated responses of cerebellar Purkinje cells
are independent of the dendritic location of granule cell synaptic inputs.
PNAS. 1994;91(11):4736–40.
3. Branco T, Hausser M. The single dendritic branch as a fundamental functional unit in the nervous system. Curr Opin Neurobiol.
2010;20(4):494–502.
P175
Dendritic morphology determines how dendrites are organized
into functional subunits
Sungho Hong1, Akira Takashima1, Erik De Schutter1
1
Computational Neuroscience Unit, Okinawa Institute of Science
and Technology Graduate University, Onna‑son, Okinawa 904‑0495,
Japan
Correspondence: Sungho Hong ‑ shhong@oist.jp
BMC Neuroscience 2016, 17(Suppl 1):P175
Studies have established that dendrites are not simple cables that
deliver synaptic inputs to a spike initiation zone in a neuron but can
also perform active transformation, which is termed “dendritic computation” [1]. In particular, it has been claimed that individual dendritic branch should function as a local computational subunit [2] and
therefore single neurons (especially pyramidal neurons) can act like
two-layer neural networks [3]. Evidence supporting these hypotheses is largely based on existence of active membrane mechanisms
in dendrites that give rise to their rich computational capabilities (e.g.
[4]) and independent operations [5]. However, dendritic morphology
BMC Neurosci 2016, 17(Suppl 1):54
is also known to play a significant role: For example, spike backpropagation is effectively prevented in the cerebellar Purkinje cells mostly
due to morphology, even when artificial active mechanisms supporting propagation are embedded in simulations [6]. Nevertheless, to our
best knowledge, there has been only few studies that quantified how
the real morphological structure can control the functional properties
of dendrites by forming subunits.
Here we address this question by combining a data-driven statistical
analysis and computational modeling approach: First, we simulated
central neurons of diverse morphological types with the passive membranes where localized inputs were injected. Response patterns in the
dendritic membrane were collected as “features” corresponding to
the input sites. Then, our dimensionality reduction/clustering procedure grouped them into clusters, which we call “subunits”. We found
that those subunits usually consist of a few nearby branches in many
neuron types, containing 2.12 ± 0.13 dendritic terminals per subunit
(mean ± SEM), whereas they consist of one or more branchlets in the
cerebellar Purkinje cells (12.9 ± 0.82 terminals). We also found that the
subunits are comparable with other functional properties such as sublinear summation of multiple synaptic inputs and spreading of a dendritic spike.
Conclusions The morphological branching pattern of a neuronal dendritic tree determines how dendrites are organized into functional
subunits. This implies that principles governing synaptic integration
and active events, such as dendritic spiking, can widely vary depending on the morphological type of the neuron.
References
1. London M, Häusser M. Dendritic computation. Annu Rev Neurosci.
2005;28:503–32.
2. Branco T, Häusser M. The single dendritic branch as a fundamental functional unit in the nervous system. Curr Opin Neurobiol. 2010;20:494–502.
3. Poirazi P, Brannon T, Mel BW. Pyramidal neuron as two-layer neural network. Neuron. 2003;37:989–99.
4. Branco T, Clark BA, Häusser M. Dendritic discrimination of temporal input
sequences in cortical neurons. Science. 2010;329:1671–5.
5. Behabadi BF, Mel BW. Mechanisms underlying subunit independence in
pyramidal neuron dendrites. Proc Natl Acad Sci USA. 2014;111:498–503.
6. Vetter P, Roth A, Häusser M. Propagation of action potentials in dendrites
depends on dendritic morphology. J Neurophysiol. 2001;85:926–37.
P176
A model of Ca2+/calmodulin‑dependent protein kinase II activity
in long term depression at Purkinje cells
Criseida Zamora1, Andrew R. Gallimore1, Erik De Schutter1
Computational Neuroscience Unit, Okinawa Institute of Science
and Technology Graduate University, Okinawa 904‑0895, Japan
Correspondence: Won Hee Lee ‑ criseida.chimal@oist.jp
BMC Neuroscience 2016, 17(Suppl 1):P176
Cerebellar long-term depression (LTD) is a form of synaptic plasticity
involved in motor learning. It is characterized as a robust and persistent decrease in the synaptic transmission between parallel fibers
(PF) and Purkinje cells (PC), which is expressed as a reduction in the
number of synaptic AMPA receptors (AMPAR). LTD signaling network
includes a PKC-ERK-cPLA2 positive feedback loop and mechanism of
AMPAR trafficking. Previous studies suggest that Ca2+/calmodulindependent protein kinase II (CaMKII) is required for the LTD induction
[1]. However, the molecular mechanism of how CaMKII contributes to
LTD is not fully understood. Noise in the signaling networks plays an
important role in cellular processes. LTD models including the CaMKII pathway have been developed [2], but they have not included the
intrinsic stochasticity of molecular interactions.
Our lab recently developed a stochastic model of the LTD signaling network including a PKC-ERK-cPLA2 feedback loop and AMPAR trafficking
[3]. In this work, we have extended the model by adding the molecular
network regulating CaMKII activity, which is known to influence LTD.
This new model was solved stochastically by STEPS (STochastic Engine
for Pathway Simulation) to simulate the influence of noise in the LTD
Page 100 of 112
signaling network [4]. Some of the most important new components
of this network include phosphatase 2A (PP2A), phosphodiesterase 1
(PDE1), cGMP/protein kinase G (PKG) and nitric oxide (NO) pathway.
Through stochastic modeling we showed that the requirement of
CaMKII activity for LTD induction is controlled by its indirect inhibition of PP2A activity, with PP2A markedly suppressing the activation
of LTD when CaMKII activity is decreased. The impairment of LTD could
be rescued by the additional PDE1 reduction when CaMKII is reduced.
In addition, the cGMP/PKG pathway supports LTD through its activation by NO. These results are congruent with previous studies of CaMKII activity [2] and make our stochastic model a potential tool to study
the effects of CaMKII, phosphatases and phosphodiesterases in LTD
molecular network.
References
1. Hansel C, de Jeu M, Belmeguenai A, Houtman SH, Buitendijk GH, Andreev
D, De Zeeuw CI, Elgersma Y. αCaMKII is essential for cerebellar LTD and
motor learning. Neuron. 2006;51:835–43.
2. Kawaguchi SY, Hirano T. Gating of long-term depression by Ca2+/calmodulin-dependent protein kinase II through enhanced cGMP signalling in
cerebellar Purkinje cells. J Physiol. 2013;591(7):1707–30.
3. Antunes G, De Schutter E. A stochastic signaling network mediates the
probabilistic induction of cerebellar long-term depression. J Neurosci.
2012;32(27):9288–300.
4. Hepburn I, Chen W, Wils S, De Schutter E. STEPS: efficient simulation of
stochastic reaction–diffusion models in realistic morphologies. BMC Syst
Biol. 2012;6:36.
P177
Reward‑modulated learning of population‑encoded vectors
for insect‑like navigation in embodied agents
Dennis Goldschmidt1, Poramate Manoonpong2, Sakyasingha Dasgupta3,4
1
Champalimaud Neuroscience Programme, Champalimaud Center
for the Unknown, Lisbon, Portugal; 2Center of Biorobotics, Mærsk
Mc‑Kinney Møller Institute, University of Southern Denmark, Odense,
Denmark; 3Riken Brain Science Institute, 2‑1 Hirosawa, Wako, Saitama,
Japan; 4IBM, IBM Research ‑ Tokyo, Tokyo, 103‑8510, Japan
Correspondence: Dennis Goldschmidt ‑ dennis.goldschmidt@neuro.
fchampalimaud.org
BMC Neuroscience 2016, 17(Suppl 1):P177
Many insects exhibit robust and efficient visual-based navigation in
complex environments [1]. Specifically, behavioral studies on ants and
bees showed that they are guided by orientation vectors based on a
process called path integration. This process allows them to estimate
their current location by integrating cues from odometry and a sunbased compass. While it is mainly applied to return back to the nest, it
also guides learning of so-called vector memories for subsequent foraging [2, 3]. Vector memories can be anchored globally to the nest or
locally to landmarks. Recent neurophysiological studies revealed that
the central complex, an insect neuropil, contains neural representations of compass [4] and odometric cues [5]. However, it is still unclear,
how these representations are involved in path integration and vector memories, and how they produce goal-directed navigation. Computational modeling has been powerful in testing hypotheses about
the underlying neural substrates and their generated behavior, and to
predict further experimental data. Previous models [6, 7] sufficiently
produced insect-like vector navigation, but they neglected biologically plausible explanations about underlying neural mechanisms that
could generate this behavior.
We present here a novel computational model of neural mechanisms
in closed-loop control for vector navigation in embodied agents. It
consists of a path integration mechanism, reward-modulated learning
of global and local vectors, random search, and action selection. The
path integration mechanism computes a vectorial representation of the
agent’s current location. The vector is encoded in the activity pattern of
circular arrays, where the angle is population-coded and the distance is
rate-coded. We apply a reward-modulated learning rule for global and
local vector memories, which associates the local food reward with the
BMC Neurosci 2016, 17(Suppl 1):54
Page 101 of 112
path integration state. A motor output is computed based on the combination of vector memories and random exploration. We show that the
modeled neural mechanisms enable robust homing and localization in
a simulated agent, even in the presence of external sensory noise. The
proposed learning rules produce goal-directed navigation and route
formation under realistic conditions. This provides an explanation for,
how view-based navigational strategies are guided by path integration.
As such, the model is the first to link behavioral observations to their
possible underlying neural substrates in insect vector navigation.
Acknowledgements: We thank Florentin Wörgötter at the Department of Computational Neuroscience in Göttingen, where most of this
work was conducted. SD acknowledges funding from the RIKEN Brain
Science Institute.
References
1. Wehner R. Desert ant navigation: how miniature brains solve complex
tasks. J Comp Physiol A. 2003;189(8):579–88.
2. Collett M, Collett TS, Bisch S, Wehner R. Local and global vectors in desert
ant navigation. Nature. 1998;394(6690):269–72.
3. Collett TS, Collett M. Route-segment odometry and its interactions with
global path-integration. J Comp Physiol A. 2015;201(6):617–30.
4. Seelig JD, Jayaraman V. Neural dynamics for landmark orientation and
angular path integration. Nature. 2015;521(7551):186–91.
5. Martin JP, Guo P, Mu L, Harley CM, Ritzmann RE. Central-complex
control of movement in the freely walking cockroach. Curr Biol.
2015;25(21):2795–803.
6. Cruse H, Wehner R. No need for a cognitive map: decentralized memory
for insect navigation. PLoS Comput Biol. 2011;7(3):e1002009.
7. Kubie JL, Fenton AA. Heading‐vector navigation based on head‐direction
cells and path integration. Hippocampus. 2009;19(5):456–79.
P178
Data‑driven neural models part II: connectivity patterns of human
seizures
Philippa J. Karoly1, Dean R. Freestone1,2, Daniel Soundry2, Levin
Kuhlmann3, Liam Paninski2, Mark Cook1
1
Department of Medicine, The University of Melbourne, Parkville VIC,
3010, Australia; 2Department of Statistics, Columbia University, New
York, NY, USA; 3Swinburne University of Technology, Hawthorn, VIC 3122,
Australia
Correspondence: Philippa J. Karoly ‑ pkaroly@student.unimelb.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P178
Here we present a model-based estimation framework for electrocorticography (ECoG) data that provides insight into mechanisms of seizures; and can be used as a clinical tool to monitor and design new
treatment strategies on a patient-specific basis.
Seizures are brief periods of abnormal, hypersynchronous neural firing that spreads across multiple cortical regions. People with epilepsy experience recurrent seizures, which are often untreatable and
of unknown cause. The data-driven estimation framework, shown in
Fig. 91, describes dynamic neural connectivity patterns during patient
seizures. Data were obtained from a clinical trial for an implantable
seizure warning device [1], which captured thousands of seizures. We
estimated mean membrane potentials and connectivity strengths
between excitatory, inhibitory and pyramidal populations using a nonlinear, assumed density filter for the neural mass equations [2].
Estimated parameters provide insights into the mechanisms of the seizure, which are not apparent from ECoG alone. For instance, Panel E
shows the seizure is preceded by a focal decrease in inhibition compared
to the surrounding channels, with widespread disinhibition during the
seizure. Joint state and parameter estimation was repeated for every seizure, and the results showed consistent, stereotypical effective connectivity patterns that differed between short (<20 s) and long seizures. This
is an important finding, as understanding the regulatory factors implicated in stopping seizures can guide new pharmaceutical treatments
and electrical counter-stimulation strategies. The successful application
of the neural mass model to study epileptic seizures supports the use of
data-driven estimation for the clinical management of epilepsy.
Fig. 91 Example estimation of a seizure recording. A Sixteen channel
electrocortiography (ECoG) of seizure (red lines indicate start and end
points). B The ECoG channels are modelled as cortical regions, each
with three coupled populations. C–G Estimation results of coupling
strength (proportional to color) between neural populations for
16 cortical regions (vertical axis), over the time span of the seizure
(horizontal axis)
References
1. Cook MJ, et al. Prediction of seizure likelihood with a long-term,
implanted seizure advisory system in patients with drug-resistant epilepsy: a first-in-man study. Lancet Neurol. 2013;12(6):563–71.
2. Freestone et al.: Data-driven neural models part I CNS 2016 abstract
submission.
P179
Data‑driven neural models part I: state and parameter estimation
Dean R. Freestone1, Philippa J. Karoly1,2, Daniel Soundry3, Levin
Kuhlmann4, Mark Cook1
1
Department of Medicine, The University of Melbourne, Parkville VIC,
3010, Australia; 2Department of Electrical and Electronic Engineering,
The University of Melbourne, Parkville VIC, 3010, Australia; 3Department
of Statistics, Columbia University, New York, NY, USA; 4Swinburne
University of Technology, Hawthorn, VIC 3122, Australia
Correspondence: Dean R. Freestone ‑ deanrf@unimelb.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P179
This work describes a novel algorithm for inferring neural activity and
effective cortical connectivity from neuroimaging data. The ability to infer
cortical network structure from data is an important step towards understanding and treating neurological disorders, such as epilepsy. However,
statistical measures for correlation in neuroimaging data are ambiguous
and bear little or no relation to physiology. On the other hand, estimating physiologically realistic connectivity is highly challenging due to the
complex, non-linear dynamics of the brain. The algorithm we present
overcomes this challenge by providing an exact solution to non-linear
inversion for a class of biologically inspired neural network models.
The presented algorithm performs joint state and parameter estimation for a class of neural model that represents interacting cortical regions as coupled nodes (shown in Fig. 92). The states of the
model represent mean cortical activity [population membrane
potentials, v(t)], and the parameters are the effective connectivity
(synaptic gain kernels, αi,e). The output voltage, vn(t), represents the
electrophysiology recording, which is inverted using a novel formulation of the Kalman filter equations for neural models [1]. The novelty of
this method is the derivation of an exact solution to the integral over
the distribution of hidden model states conditioned on previous data.
We provide results showing that the new algorithm demonstrates
higher estimation accuracy and greater computational tractability than existing inference methods for neural models. We also show
example estimation results from an electrical recording of a human
seizure (shown in Fig. 92). This new method for data-driven inference
represents an important contribution to online diagnostic applications, in particular for the treatment of epilepsy [2].
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 92 Data-driven model estimation. A The basic unit of a neural
model is described by the mean membrane potential, vn(t), of a
neural ensemble and synaptic inputs. B, C Pre-synaptic firing rates are
convolved with the excitatory/inhibitory kernel to generate membrane potential fluctuations. D The resulting membrane potential is
converted to an output firing rate via a sigmoidal transform. E Electrical recording of a seizure. F Estimated gain parameters during seizure
References
1. Freestone DR, et al. Estimation of effective connectivity via data-driven
neural modeling. Front Neurosci. 2014;8:383
2. Karoly, et al. Data-driven neural models part II: connectivity patterns of
human seizures. CNS 2016 abstract.
P180
Spectral and spatial information processing in human auditory
streaming
Jaejin Lee1, Yonatan I. Fishman2, Yale E. Cohen1
1
Department of Otorhinolaryngology – Head and Neck Surgery,
University of Pennsylvania, Philadelphia, PA 19104, USA; 2Department
of Neurology, Albert Einstein College of Medicine, Bronx, NY 10461, USA
Correspondence: Jaejin Lee ‑ jaejin@mail.med.upenn.edu
BMC Neuroscience 2016, 17(Suppl 1):P180
The purpose of auditory system is to transform acoustic stimuli from
the external environment to sound perception. To achieve this goal,
the auditory system needs to analyze a mixture of stimuli that originate from independent sources and distinguish individual sound
sources in the auditory scene. It is believed that the auditory system
groups and segregates auditory stimuli based on their regularities,
but the neural basis of how regularities relate to sound perception
is not well known. The ventral pathway in the brain is involved in
auditory perception whereas the dorsal pathway is involved in spatial processing and audiometer processing. We are interested in how
the spatial information is represented in the ventral pathway during
perceptual auditory streaming tasks that use spatial information.
We first developed a novel task based on [1] in which human listeners can segregate streams using spectral or spatial information and
detect the deviant tone. An array of 13 free field speakers with different spatial distributions were used to play the stimuli. The frequency
difference between streams and the spatial separation were varied to
explore how the spectral and spatial information interplay in the auditory streaming task. We also manipulated other acoustic features of
the stimuli to understand how different acoustic cues can affect the
auditory streaming performance. We found that the ability to segregate the streams is vastly improved when there is spatial information
available in addition to spectral information. Also, we further analyzed
the behavioral data to get psychophysical kernels and fit the data to
variants of sequential sampling models related to the drift diffusion
model(DDM) [2] to quantify the effects of sequence coherence on the
decision making process.
References
1. Sussman E, Steinschneider M. Attention effects on auditory scene analysis in children. Neuropsychologia. 2009;47:771–85.
2. Liu A, Tsunada J, Gold J, Cohen Y. Temporal integration of auditory information is invariant to temporal grouping cues. eNeuro.
2015;2(2):e0077-14.
Page 102 of 112
P181
A tuning curve for the global effects of local perturbations
in neural activity: mapping the systems‑level susceptibility of the
brain
Leonardo L. Gollo1, James A. Roberts1, Luca Cocchi1
1
Systems Neuroscience Group, QIMR Berghofer Medical Research
Institute, Herston, QLD 4006, Australia
Correspondence: Leonardo L. Gollo ‑ leonardo.l.gollo@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P182
The activity of the human brain in a state of rest exhibits a defined pattern of functional connectivity, and a small set of functional networks,
which comprise regions that are highly correlated and are mostly distinctive from one another [1]. However, despite recent efforts [2, 3], the
effects of local perturbations into endogenous whole brain dynamics are not yet clearly understood [4]. To gain insights into the global
effects of a focal perturbation, we simulate the human brain dynamics
using a weighted high-resolution connectome of 513 cortical regions
[5]. The cortical dynamics is modelled by a canonical oscillatory model,
introducing heterogeneous dynamics between cortical regions as a
function of the anatomical nodal strength (sum of weights). Such heterogeneity leads to a hierarchy of time scales of cortical regions recapitulating the known anatomical hierarchy, with peripheral regions
having fast time scales and core regions with slow time scales [6].
Results showed that nodal diversity is not just a crucial element to
improve the model’s performance [7, 8], but also to reproduce the
experimental data of variations in functional connectivity following
local inhibitory transcranial magnetic stimulation (TMS). We find a
large variation in the overall effect of functional connectivity following local stimulation. Specifically, the inhibition of hub nodes causes
increased anti-correlated activity, whereas inhibition of peripheral
nodes caused increased correlated activity with the rest of the brain.
The intensities of the variations in functional connectivity with respect
to baseline were also highly variable and stronger for intermediary
nodes that were not hubs or peripheral regions. Moreover, depending
on the weights of the cortical regions, changes in functional connectivity form a tuning curve (Fig. 93). Overall, our findings suggest a key
role of local temporal dynamics to explain the widespread effects of
focal perturbations in neural activity.
References
1. Power JD, Cohen AL, Nelson SM, Wig GS, Barnes KA, Church JA, Vogel AC,
Laumann TO, Miezin FM, Schlaggar BL, Petersen SE. Functional network
organization of the human brain. Neuron. 2011;72:665–78.
Fig. 93 Changes in functional connectivity with respect to baseline
after inhibitory stimulation as a function of cortical weight of the
structural connectivity matrix. Red line: mean uniform bins curve
smoothed; dashed line: mean weight
BMC Neurosci 2016, 17(Suppl 1):54
2.
3.
4.
5.
6.
7.
8.
Cocchi L, Sale MV, Lord A, Zalesky A, Breakspear M, Mattingley JB. Dissociable effects of local inhibitory and excitatory theta-burst stimulation
on large-scale brain dynamics. J Neurophysiol. 2015;113:3375–85.
Kunze, T, Hunold A, Haueisen J, Jirsa V, Spiegler A. Transcranial direct
current stimulation changes resting state functional connectivity: a
large-scale brain network modeling study. NeuroImage. doi:10.1016/j.
neuroimage.2016.02.015.
Sale MV, Mattingley JB, Zalesky A, Cocchi L. Imaging human brain networks to improve the clinical efficacy of non-invasive brain stimulation.
Neurosci Biobehav Rev. 2015;57:187–98.
Roberts JA, Perry A, Lord AR, Roberts G, Mitchell PB, Smith RE, Calamante
F, Breakspear M. The contribution of geometry to the human connectome. NeuroImage. 2016;124:379–93.
Gollo LL, Zalesky A, Hutchison RM, van den Heuvel M, Breakspear M.
Dwelling quietly in the rich club: Brain network determinants of slow
cortical fluctuations. Philos Trans R Soc B Biol Sci. 2015;370:20140165.
Mejias JF, Longtin A. Optimal heterogeneity for coding in spiking neural
networks. Phys Rev Lett. 2012;108:228102.
Gollo LL, Copelli M, Roberts JA. Diversity improves performance in excitable networks. arXiv preprint arXiv:1507.05249. 2015 Jul 19.
P182
Diverse homeostatic responses to visual deprivation mediated
by neural ensembles
Yann Sweeney1, Claudia Clopath1
1
Department of Bioengineering, Imperial College London, UK
Correspondence: Yann Sweeney ‑ y.sweeney@imperial.ac.uk
BMC Neuroscience 2016, 17(Suppl 1):P182
Visual deprivation paradigms provide crucial insight into the homeostatic response in visual cortex. We explore how neurons within functional ensembles may exhibit correlated homeostatic responses to
visual deprivation, and how the source of common inputs to these
ensembles determine the extent of their homeostatic recovery. We
hypothesise that common inputs from non-visual stimuli are responsible for driving recovery from visual deprivation.
We simulate development during spontaneous and evoked activity
in a recurrent network model of visual cortex in which Hebbian and
homeostatic synaptic plasticity is implemented. This leads to the emergence of highly interconnected ensembles of neurons driven by either
common visual or common non-visual inputs. When we then deprive
the developed network of visual input, the homeostatic response is a
strengthening of activity within ensembles which share common nonvisual inputs. A broad reduction in inhibition across the network is also
observed. Interestingly, the magnitude of the homeostatic response
depends on the size of these ensembles, with larger ensembles more
likely to fully recover from visual deprivation. Our results demonstrate
the importance of investigating functional plasticity of ensembles triggered by sensory deprivation paradigms.
Acknowledgements: This research was supported by the Engineering
and Physical Sciences Research Council (EPSRC), the Leverhulme Trust
and Google Faculty Award.
P183
Opto‑EEG: a novel method for investigating functional
connectome in mouse brain based on optogenetics and high
density electroencephalography
Soohyun Lee1,3, Woo‑Sung Jung1,2, Jee Hyun Choi3
1
Department of Physics, POSTECH, Pohang, 37673, South Korea;
2
Department of Industrial and Management Engineering, POSTECH,
Pohang, 37673, South Korea; 3Center for Neuroscience, KIST, Seoul, 02792,
South Korea
Correspondence: Jee Hyun Choi ‑ jeechoi@kist.re.kr
BMC Neuroscience 2016, 17(Suppl 1):P183
Connectome, comprehensive structural description of the network of
elements and connections forming the brain [1], is fundamental for
understanding the brain functions. Recent advances in optical imaging techniques allow us to be feasible to structural connectivity. But
differently from structural connectivity, the functional connectivity is
Page 103 of 112
Fig. 94 Propagation patterns of optical stimulation at each frequency in thalamocotical circuit. Beta frequency stimulus propagated
S1–M1 strongly, but gamma frequency case, contralateral propagation is dominant. Blue bars indicate optical stimulus in left VPM
altered by condition such as brain states, input types and pathological conditions. To construct functional connectome, the techniques
to map individual functional circuit and control specific neuronal
activity have been needed. However, the current functional brain
mapping techniques have limitations to obtain the map of the functionally correlated brain activity in freely moving mouse model. Here,
we introduce novel functional brain mapping technique for mouse
model by high density electroencephalography [2] under optogenetic
stimulus, which we referred as opto-EEG. Opto-EEG tool enables us to
investigate the functionally connected neuronal circuit with high spatial and temporal resolution. We stimulated ventral posterioromedial
thalamic nucleus (VPM) with various frequencies for verifying different frequency dependency of functional connectome. Stimulation
of VPM induced sequential activations of ipsilateral somatosensory
cortex (S1) followed by ipsilateral motor cortex (M1), contralateral M1
and contralateral S1. The power based analysis result showed information flow between S1 and M1 was maximized under beta frequency
stimulus. On the other hand, latency-based result showed minimized
interhemispheric transfer latency under gamma band (Fig. 94). This
example indicates that opto-EEG makes it possible to be used to characterize the functional connectivity under temporally precise control
of specific neuronal circuits, provide new insights into brain exploration capabilities of functional connectome, and be applied to discover
neuromodulation method for treatment of disease or pathologies.
References
1. Sporns O, Tononi G, Kotter R. The human connectome: A structural
description of the human brain. PLoS Comput Biol. 2005;1(4):e42.
2. Lee M, Kim D, Shin HS, Sung HG, Choi JH. High-density EEG recordings of
the freely moving mice using polyimide-based microelectrode. J Vis Exp.
2011;(47).
P184
Biphasic responses of frontal gamma network to repetitive sleep
deprivation during REM sleep
Bowon Kim1,2, Youngsoo Kim3, Eunjin Hwang1, Jee Hyun Choi1,2
1
Center for neuroscience, Korea Institute of Science and Technology,
Seoul, South Korea; 2Department of Neuroscience, University of Science
and Technology, Daejon, South Korea; 3Department of Psychiatry, VA
Boston Healthcare System & Harvard Medical School, Brockton, MA, USA
Correspondence: Jee Hyun Choi ‑ jeechoi@kist.re.kr
BMC Neuroscience 2016, 17(Suppl 1):P184
Prefrontal cortex has been known to be less activated [1] and decoupled from the other cortical area in REM sleep [2]. In our previous study
of chronic sleep deprivation (SD) in mice model, we observed that the
5 successive days of SD (SD 1–5, 18 h sleep deprivation in each day)
BMC Neurosci 2016, 17(Suppl 1):54
Fig. 95 The pairs with statistically significantly increased (red) or
decreased (blue) PSI of gamma oscillations (Student t test, p < 0.05).
Only the pairs from the prefrontal cortex were depicted here. SD and
R stand for sleep deprivation and recovery days, respectively
induced a monotonic increase of the prefrontal gamma oscillation (30–
40 Hz) in REM sleep as the sleep pressure increased. However, the functional role of this increased gamma oscillation was not answered. Here,
we investigated the functionality of the increased prefrontal gamma
in sleep deprived nights by calculating the connectivity between the
prefrontal cortex and the other cortical regions [3]. Phase synchrony
index (PSI) was employed to minimize the volume conduction in high
density EEG microarray. In the first day of the sleep deprivation (SD 1),
we observed statistically significant increases in gamma connectivity within the bilateral prefrontal regions and between prefrontal and
ipsilateral somatosensory cortex. However, as the sleep deprivation
continued, an opposite response of prefrontal-somatosensory gamma
connectivity was observed in a way that the PSI between these
two areas become insignificant in SD 3 and statistically significantly
decreased in the SD 5, which remained even after the 3rd day of recovery after the sleep deprivation (R 3). The area of decreased gamma
connectivity became broader as well. On the other hand, the intracortical connectivity within prefrontal connectivity remained elevated
throughout the sleep deprivation and recovery days. This result implies
that the increased prefrontal gamma oscillation due to the homeostatic
response of REM sleep does not participate in the information transfer
from prefrontal to the other cortical area (Fig. 95).
Acknowledgements: This research was supported by the Korean
National Research Council of Science and Technology (No.
CRC-15-04-KIST).
References
1. Maquet P, et al. Functional neuroanatomy of human rapid-eye-movement sleep and dreaming. Nature. 1996;383(6596):163–6.
2. Castro S, et al. Coherent neocortical 40‐Hz oscillations are not present
during REM sleep. Eur J Neurosci. 2013;37(8):1330–9.
3. Hwang E, McNally JM, Choi JH. Reduction in cortical gamma synchrony
during depolarized state of slow wave activity in mice. Front Syst Neurosci. 2013;7.
P185
Brain‑state correlate and cortical connectivity for frontal
gamma oscillations in top‑down fashion assessed by auditory
steady‑state response
Younginha Jung1,2, Eunjin Hwang1, Yoon‑Kyu Song2, Jee Hyun Choi1,3
1
Center for Neuroscience, Korea Institute of Science and Technology,
Seoul 02792, Korea; 2Program in Nano Science and Technology, Seoul
National University, Seoul 08826, Korea; 3Department of Neuroscience,
University of Science and Technology, Daejon 34113, Korea
Correspondence: Jee Hyun Choi ‑ jeechoi@kist.re.kr
BMC Neuroscience 2016, 17(Suppl 1):P185
Cortical gamma rhythm, particularly in the frequency range of
30–50 Hz, has received intensive attention as neural correlates of cognitive process [1]. On the other hand, diminished cognitive flexibility,
one of the typical symptoms in psychiatric disorders, is closely associated with disturbances in neural oscillations, specifically gamma
band [2]. To quantify gamma-band oscillation, auditory steady-state
Page 104 of 112
response (ASSR) evoked by repetitive auditory stimulus given at a rate
of 40 Hz has been used as a prominent approach which reflects neural
efficiency for maintaining gamma oscillation [3]. Despite its diagnostic
advantages, there is less discussion whether ASSRs are modulated by
endogenous top-down effect. The present research attempts to investigate top-down influences on ASSR by analyzing in vivo mouse data.
Experimental data in this study were obtained from 38-channel
mouse epidural electroencephalogram during auditory steady-state
stimulus. Interestingly, there were two distinctive topographic maps
of EEG spectral power and the notable difference between topographies was the presence or absence of frontal responses. By comparing topographic results, we hypothesized that frontal ASSRs reflect
top-down functioning. The analytic approaches taken in this work
are based on brain-state alteration and regional connectivity. The
first research question in the data analysis is that frontal ASSRs switch
states of arousal via top-down control. Video-based behavior analysis
was adapted to classify arousal states into wakefulness and drowsiness and the proportion of arousal behavior in two topographies
were determined. In addition, comparison of delta spectral power
for topographic patterns could explain frontal engaged sleep state
modulation. The second study question is about early stages of auditory ascending pathway in each topographic pattern. Magnitude and
latency of auditory evoked potentials and gamma spectral power
were analyzed in inferior colliculus and primary auditory cortex. The
third question in data analysis is functional connectivity among cortical regions and, in detail, phase-locking value and directed phase lag
index were calculated in frontotemporal and inter-frontal coupling.
Overall, the current results show that frontal lobe contributes substantially to ASSR and imply that it is important to consider the frontal involvement in auditory steady-state signal processing. Together,
these methodologies could provide important insights to clinical
research by demonstrating top-down modulation. Investigating
gamma oscillatory activity across cortical regions potentially provides
deeper understanding for dysfunction in neurological disorders and
furthermore gives clues to determine neural circuit disruption.
References
1. Fries P, Reynolds JH, Rorie AE, Desimone R: Modulation of oscillatory neuronal synchronization by selective visual attention. Science.
2001;291(5508):1560–3.
2. Kwon JS, O’Donnell BF, Wallenstein GV, Greene RW, Hirayasu Y, Nestor PG,
Hasselmo ME, Potts GF, Shenton ME, McCarley RW. Gamma frequencyrange abnormalities to auditory stimulation in schizophrenia. Arch Gen
Psychiatry. 1999;56(11):1001–5.
3. Picton TW, John MS, Dimitrijevic A, Purcell D. Human auditory steadystate responses. Int J Audiol. 2003;42(4):177–219.
P186
Neural field model of localized orientation selective activation
in V1
James Rankin1, Frédéric Chavane2
1
Center for Neural Science, New York University, 4 Washington Place,
10003 New York, NY, USA; 2Institut de Neuroscienes de la Timone (INT),
CNRS & Aix‑Marseille University, 27 Boulevard Jean Moulin, 13005
Marseille, France
Correspondence: James Rankin ‑ james.rankin@nyu.edu
BMC Neuroscience 2016, 17(Suppl 1):P186
Voltage imaging experiments in primary visual cortex [1] have shown
that local, oriented visual stimuli elicit stable orientation-selective
activation within the stimulus retinotopic footprint. The cortical activation dynamically extends far beyond the retinotopic footprint, but
the peripheral spread stays non-selective—a surprising finding given
a number of studies showing the orientation specificity of longrange connections, e.g. [2]. We study the dynamics of these inputdriven localized states in a planar neural field model building on an
earlier theoretical study using radially symmetric inputs [3]. Here we
use a new anatomically-motivated connectivity profile and extend
the model to multiple sub-populations encoding orientation. For
canonical choices of connectivity profile (such as a radial difference of
BMC Neurosci 2016, 17(Suppl 1):54
Page 105 of 112
Gaussians), localized orientation selectivity arises. However, unlike the
experimental observations, the selective activation is unstable during transient dynamics. In the new connectivity profile defined in our
study, the range of excitatory and inhibitory connections and the orientation selectivity of those connections are controlled with separate
parameters. We demonstrate how peaks in the number of excitatory
connections at each hyper-column distance [4] are crucial in stabilizing the transient, local orientation selective activation. If these peaks
in excitation are non-realistically exaggerated, we demonstrate that
spurious selectivity (not matching preference map) could arise in the
peripheral spread. Furthermore, although orientation selectivity of
connections increases accuracy of the selective activation within the
retinotopic footprint, it can also lead to orientation selective activation
in the periphery. Our parameter exploration shows that with a balance
in the sharpness of peaks in long-range excitatory connections and
the selectivity of these connections, we can capture the correct localized selective activation, the non-selective peripheral spread and the
stable transient dynamics.
Conclusions Typical choices of connectivity profile in planar models
of cortex fail to produce important aspects of the observed cortical
spread of activation. We developed a more realistic connectivity profile
inspired by anatomical data that, used in conjunction with our planar
multiple sub-population model, captures all key spatial and temporal aspects of the cortical spread of activation. For the first time, our
study shows that the unexpected experimental findings of [1] can be
accounted for with a realistic balance between the sharpness of peaks
in long-range excitation and orientation selectivity of connections.
References
1. Chavane F, Sharon D, Jancke D, Marre O, Frégnac Y, Grinvald A. Lateral
spread of orientation selectivity in V1 is controlled by intracortical cooperativity. Front Syst Neurosci. 2011;5:4
2. Bosking WH, Zhang Y, Schofield B, Fitzpatrick D. Orientation selectivity
and the arrangement of horizontal connections in tree shrew striate
cortex. J Neurosci. 1997;17:2112–27.
3. Rankin J, Avitabile D, Baladron J, Faye G, Lloyd DJ. Continuation of localized coherent structures in nonlocal neural field equations. SIAM J Sci
Comput. 2014;36:B70–93.
4. Buzás P, Eysel U, Adorján P, Kisvárday Z. Axonal topography of cortical
basket cells in relation to orientation, direction, and ocular dominance
maps. J Comp Neurol. 2001;437:259–85.
P187
An oscillatory network model of Head direction and Grid cells
using locomotor inputs
Karthik Soman1, Vignesh Muralidharan 1, V. Srinivasa Chakravarthy1
1
Department of Biotechnology, Indian Institute of Technology Madras,
Chennai, Tamil Nadu, India
Correspondence: V. Srinivasa Chakravarthy ‑ schakra@iitm.ac.in
BMC Neuroscience 2016, 17(Suppl 1):P187
The model (Fig. 96A) takes proprioceptive inputs coming from the
joint angles of the two limbs. Locomotor rhythms are modeled as two
sinusoidal oscillators whose amplitudes are modulated by the curvature and the speed of simulated animal. These inputs are gated using a
leaky integrate and fire (LIF) neuron that spikes at a fixed frequency so
that the curvature and speed information from respective limb oscillations are extracted out and given to two oscillatory neural networks
separately. The oscillatory neural networks are modeled as Kuramoto
networks in which the phase is modulated by the curvature of the
path traversed by the simulated animal. Synchrony between the two
clusters of oscillators is quantified in terms of phase coherence and
phase difference. The synchrony parameters are further used to train
a one dimensional self organizing map (SOM), whose neurons display
head direction-tuned responses. Each HD response is transformed to
a cosine response which is further given to path integration (PI) layer.
PI layer is again a network of Kuramoto oscillators whose phase is integrated as a function of HD responses. PCA is done on the PI values. The
top few principal components (PC) corresponding to the largest Eigen
values are rearranged in increasing order and taken as the weight
Fig. 96 A The model architecture. B Hexagonal firing field of a single
neuron in the outer layer of the model while remaping its response
on the visual space
connections from the PI layer to an outer 1-D layer of neurons. While
remapping the neural response of the third and fourth PCs, square
grid fields were observed while the fifth and sixth PCs gave
We present a model of head direction (HD) and grid cells formed
purely from idiothetic (locomotor) inputs. Grid cells are a class of spatial cells located in the medial Entorhinal Cortex which is assumed to
perform path integration and characterized by its unique hexagonal
firing fields [1]. Empirically it is proven that HD cells, another class of
spatial cells which encode the heading direction of an animal, form
the major input to the grid cell. Existing computational models of grid
cells make artificial assumptions like existence of HD cells with a phase
differences that are integral multiples of 60° [2]. The aim of the study
is to model grid cell firing without imposing these special constraints.
Hexagonal grid fields (Fig. 96B). Further analysis showed that PCs were
sinusoidal vectors. Investigation of the correlation values between the
adjacent rows of the covariance matrix of PI pointed out its similarity to circulant matrices. This was in corroboration with the circulant
matrix theorem which states that a circulant matrix of any size gives
rise to sinusoidal Eigen vectors. Hence this is a generalized model
which provides a theoretical basis for the formation of both hexagonal
and non hexagonal grid fields and possibly other spatial cells which
are actually projections of the PI values onto sinusoidal orthonormal
basis vectors.
References
1. Hafting T, Fyhn M, Molden S, Moser M-B, Moser EI. Microstructure of a
spatial map in the entorhinal cortex. Nature. 2005;436(7052):801–6.
2. Burgess N, Barry C, O’Keefe J. An oscillatory interference model of grid cell
firing. Hippocampus. 2007;17(9):801–12.
BMC Neurosci 2016, 17(Suppl 1):54
P188
A computational model of hippocampus inspired by the
functional architecture of basal ganglia
Karthik Soman1,*, Vignesh Muralidharan1,*, V. Srinivasa Chakravarthy1
1
Department of Biotechnology, Indian Institute of Technology Madras,
Chennai, Tamil Nadu‑ 600036, India
Correspondence: V. Srinivasa Chakravarthy ‑ schakra@iitm.ac.in
BMC Neuroscience 2016, 17(Suppl 1):P188
* Both authors have equal contribution.
We present a networkmodel of hippocampus (HC) inspired by the
functional architecture of the basal ganglia (BG). The model describes
the role of hippocampus in spatial navigation and is cast in reinforcement learning (RL) framework (Fig. 97A). There is a corpus of literature which states that hippocampus is a key player in spatial learning
because of the enriched sensory information that arrives at the portals of HC, the entorhinal cortex (EC), from the sensory cortical areas
[1]. The model simulates the Morris water maze task wherein a virtual
agent navigates inside a circular pool to find an invisible platform
using the spatial context from the environment.
In order to model the ability of HC to learn spatial context, we simulate a circular pool surrounded by six distinguishable poles of equal
heights. As the agent/animal navigates, the size of retinal image
of each pole varies with distance between the agent and the pole.
Reward is given to the agent when it reaches the platform. The
abstract form of the visual input is given to EC which has afferent and
efferent projections from ventral tegmental area (VTA), one of the
dopamine centers in the mid brain. Temporal difference (TD) error
generated from VTA is used to update the synaptic weights for value
computation of the sensory input in EC. Additionally an action vector
defining the direction of the agent’s next step also forms a feedback
input to the EC. We describe the functional anatomy of HC in terms
of two pathways: a direct pathway between EC and CA1 and an indirect pathway between EC and CA1 via dentate gyrus (DG), and CA3.
A quantity known as Value difference, that represents afferent dopamine signals in EC, is thought to control switching between these
pathways. Desynchronized activity generated by the DG–CA3 loop in
the indirect pathway aids the agent to explore the space. Direct pathway facilitates the agent’s navigation. The difference in the responses
from these two pathways is computed in CA1 and is relayed to subiculum (Sbc) which computes the direction of the next step. Output
Fig. 97 The model architecture (A) used to simulate the water maze
task indicating the notion of a direct and an indirect pathway. The
value function (B) developed after training the agent for 10 trials, the
value peaks near to the platform. The escape latency (C) through trials
shows that the agent has learnt the task with increased hippocampal
dependence in the earlier stages and cortical dependence in the later
stages of learning. The spectrogram (D) of the activity of CA3 as the
function of time shows desynchronization while active exploration
of the maze (0–45 s) and synchronized activity upon reaching the
platform (45–65 s)
Page 106 of 112
of Sbc is communicated to higher motor areas (MC), modeled as a 1-D
Continuous attractor neural network, via deeper layers of EC. MC also
receives direct inputs from sensory areas so that the output of MC is
the weighted sum of responses from the sensory cortical areas and HC
respectively. MC response is used to update the next step. Cortico-cortical pathway (CCP) connections are updated using the TD error as well
as the velocity generated from HC as a target. As the value function
matures (Fig. 97B), contribution from HC declines; the CCP connections gain the upper hand and the agent reaches the platform faster
(Fig. 97C). Thus, after training, the CCP can drive navigation without
the involvement of HC. Analysis of CA3 activity shows desynchronization during active exploration and synchronizationupon reaching the
platform (Fig. 97D). This resonates with experimental results suggesting that low-amplitude theta waves correspond to desynchronized
activity during exploration, whereas the sharp waves during nonexploratory states correspond to synchronized activity [2].
References
1. Sukumar D, Rengaswamy M, Chakravarthy VS. Modeling the contributions of Basal ganglia and Hippocampus to spatial navigation using
reinforcement learning. PloS One. 2012;7(10):e47467.
2. Buzsáki G. Theta oscillations in the hippocampus. Neuron.
2002;33(3):325–40.
P189
A computational architecture to model the microanatomy of the
striatum and its functional properties
Sabyasachi Shivkumar1, Vignesh Muralidharan1, V. Srinivasa Chakravarthy1
1
Department of Biotechnology, Indian Institute of Technology Madras,
Chennai, Tamil Nadu, India‑600036
Correspondence: V. Srinivasa Chakravarthy ‑ schakra@iitm.ac.in
BMC Neuroscience 2016, 17(Suppl 1):P189
We propose a computational model of the functional architecture of
the striatum. Anatomical and physiological evidence suggests that
the microstructure of the striatum maps the sensory-motor information from the cortex in complex patterns. The dorsal striatum can be
differentiated into centre surround regions called striosomes and
matrisomes [1]. In the proposed striatum model, striosomes map the
state space and the matrisomes map the action space. The model consists of a hierarchical two-level self organizing map (SOM), wherein
the higher level SOM is trained on the state values and a sub-SOM
layer containing multiple smaller SOMs are trained on action values.
Neurons of ‘action SOMs’ are activated by neurons of ‘state SOMs’ The
scheme of mapping of state space and action space onto the proposed architecture is given in Fig. 98A where the red area represents
the striosomes and green area represents the matrisomes. We have
also shown previously that such feature representation in a SOM
layer can be used to develop value functions for sensory state spaces
[2]. Thus to compute the state and action values, the activities of the
respective SOMs were mapped to individual neurons by state and
action weight vectors respectively. These weights were trained by the
Fig. 98 A Centre surround mapping in striosomes and matrisomes.
The striosomes highlighted by red, map the states and the matrisomes highlighted by green, map the actions. B Value function
map in the multiple context setting where the reward is present at
the top left corner and the bottom right corner. C Switching of the
modules based on the environmental contexts. The reward changes
every 1000 episodes and the corresponding change in module with
episode is shown
BMC Neurosci 2016, 17(Suppl 1):54
temporal difference error which represents the dopamine signals from
the Substantia Nigra pars compacta (SNc) based on the reward from
the environment. Action selection was performed by using the action
values with exploration.
The model was further extended to reflect striatal modularity, which
could also be exploited to solve modular RL tasks with varying contexts [3]. This is done by using multi-SOMs, where multiple SOMs
compete with each other to represent the input space. Biologically,
this competition between different local striatal maps can be thought
to be carried out by striatal interneurons. Using the above described
striatal model as a single module, multiple modules were created. The
higher level SOMs in these modules generate a responsibility signal,
which represent the ability of the module to best represent that context. This was used to select the modules, a selection process which is
probably carried out by the tonically active neurons (TANs). To validate
this overall architecture, we tested this on the gridworld problem with
a 10 × 10 grid and 4 actions. The reward is present at the corner of
the grid in the first case and in subsequent case of modular RL framework, the reward is placed at one of two opposite corners. The value
function map is built between the two modules and their switching at
regular intervals is given in Fig. 98B, C respectively.
References
1. Graybiel A, Flaherty A, Gimenez-Amaya J-M. Striosomes and matrisomes.
In: The basal ganglia III edn. Berlin: Springer; 1991. p. 3–12.
2. Krishnan R, Ratnadurai S, Subramanian D, Chakravarthy VS, Rengaswamy
M. Modeling the role of basal ganglia in saccade generation: Is the indirect pathway the explorer? Neural Networks. 2011;24(8):801–13.
3. Amemori K-I, Gibb LG, Graybiel AM. Shifting responsibly: the importance
of striatal modularity to reinforcement learning in uncertain environments; 2011.
P190
A scalable cortico‑basal ganglia model to understand the neural
dynamics of targeted reaching
Vignesh Muralidharan1, Alekhya Mandali1, B. Pragathi Priyadharsini1, Hima
Mehta1, V. Srinivasa Chakravarthy1
1
Department of Biotechnology, Indian Institute of Technology Madras,
Chennai, Tamil Nadu‑600036, India
Correspondence: V. Srinivasa Chakravarthy ‑ schakra@iitm.ac.in
BMC Neuroscience 2016, 17(Suppl 1):P190
We present a scalable network model of the basal ganglia (BG) to highlight its role in performing simple reaching movements. The model
consists of the following components: a 2-joint arm model (AM), a
layer of motor-neurons in the spinal cord (MN), the proprioceptive cortex (PC), the motor cortex (MC), the prefrontal cortex (PFC) and the BG
(Fig. 99A). The arm model has two joints each consisting of an agonist
and an antagonist muscle pair innervated by a pair of motor neurons;
the muscles in turn control the position of the arm in 2D space. The PC
receives information about the muscle length and tension, thought to
be originating from muscle spindles and Golgi tendon organs of the
muscle. The MC then uses the sensory map information from the PC to
develop a motor map of the arm. The MC activity is also modulated by
the BG which uses reward information to make the arm learn to reach
the target. The MC then sends these signals to respective muscles of
the arm via the motor neurons (MN) to perform the movement. Since
the existence of maps has been well established in the cortex, the sensory map of the PC, and the map from PC to MC were modelled using
the self-organizing map (SOM) algorithm [1]. The motor command is
thought to arise from the PFC, which specifies the goal to be reached.
The MC therefore combines inputs from three sources: the PC, the prefrontal cortex (PFC), and the BG (from GPi via the thalamus). To enable
this summation dynamically, MC was modelled as a continuous attractor neural network (CANN), wherein stable activity in CANN space corresponds to an equilibrium position of the arm in the workspace.
Training of the model proceeds as follows. A target is chosen by activating corresponding neurons in the PFC. The arm makes exploratory
movements driven by the Indirect Pathway of BG and gets rewarded
when it reaches the target. Now BG uses this reward information and
Page 107 of 112
Fig. 99 The model architecture (A) with the different modules
aiding in reaching movements. The comparison of controls and PD’s
approach to a target (B) and the appearance of PD symtoms including tremor and rigidity as a function of distance to the target
the corresponding arm position to transform it into a value profile over
the arm’s workspace such that the trained value peaks at the target
positions. As the training of model proceeds, the arm reaches the goal
position faster and faster as BG stochastically climbs over the trained
value function [2]. Furthermore, the connections from PFC and MC
are also trained on successful reach, so that the motor command can
directly activate the motor cortex thereby producing rapid movement
avoiding the slow search conducted by the BG. The model exhibits
all stages of motor learning i.e., slow movements dominated by the
BG during early stages and cortically driven fast movements at later
stages. The simulation results show PD symptoms like tremor which
could be attributed to synchronized oscillations in STN-GPe (Fig. 99B).
References
1. Chen Y, Reggia JA. Alignment of coexisting cortical maps in a motor
control model. Neural Comput. 1996;8(4):731–55.
2. Magdoom K, Subramanian D, Chakravarthy VS, Ravindran B, Amari S-I,
Meenakshisundaram N. Modeling basal ganglia for understanding parkinsonian reaching movements. Neural Comput. 2011;23(2):477–516.
P191
Emergence of radial orientation selectivity from synaptic
plasticity
Catherine E. Davey1, David B. Grayden1,2, Anthony N. Burkitt1
1
Department of Electrical and Electronic Engineering, University
of Melbourne, Victoria, 3010, Australia; 2Centre for Neural Engineering,
University of Melbourne, Victoria, 3010, Australia
Correspondence: Catherine E. Davey ‑ cedavey@unimelb.edu.au
BMC Neuroscience 2016, 17(Suppl 1):P191
The ability to learn and recall are primary functions of the brain. Synaptic plasticity is one of the key mechanisms by which we learn and
adapt to our environment, and describes the process by which neuronal connection strengths are modified in response to environmental inputs [1]. There has been significant research effort invested into
identifying the general principles of plasticity in neural networks,
in order to garner insight into the learning process. The ability for
animals to see and hear prior to birth is evidence of learning before
exposure to ongoing external sensory signals. Consequently, cortical
structure can be created, to some extent, in the absence of structured
input. In a three-paper series, Linsker outlined a process by which cortical learning may occur prior to birth [2–4]. Linsker’s model identified
particular spatial distributions of synaptic connectivity that are sufficient to induce the development of circularly symmetric cells in a system driven only by noisy input [2]. Furthermore, Linsker [3] revealed
that orientation selective cells may develop by the sixth layer of processing. However, the resulting preferred orientation was a random
function of stochastic weight initialisations [5].
Radial selectivity describes a tendency for cells to have the preferred
orientation biased towards a central point, and has been observed
in several cortical structures, including the visual cortex [6] and the
auditory cortex [7]. In this study we reveal the mechanism by which
radial orientation selectivity can emerge from synaptic plasticity in
the absence of structured input. Linsker’s model assumed that cells
BMC Neurosci 2016, 17(Suppl 1):54
within a laminar had an identical distribution of synaptic connection
densities. This assumption is modified in this study to allow synaptic
connection densities to change as a function of a cell’s radial distance
to the centre of the laminar. The proposed network provides for spatially larger receptive fields as cells become progressively distal in the
laminar, which is in keeping with electrophysiological and anatomical results. We show, both analytically and computationally, that this
slightly modified network prompts the evolution of orientation selective cells with a predictable radial preference, in the third layer of neural processing. Importantly, this proposal maintains Linsker’s intent for
a minimal set of model assumptions, ensuring that the resulting structure is robust to details and parameter values of the model used, and
that general principles of plasticity are established. Consequently, our
results are applicable to cortical learning generally. The mechanisms
developed in this study could play a central role in the development of
radial orientation selectivity in the visual cortex.
Acknowledgements: This research was supported under Australian
Research Council’s Discovery Projects funding scheme (Project Number DP140102947).
References
1. Hughes JR. Post-tetanic potentiation. Phys Rev. 1958;38(1):91–113.
2. Linsker R. From basic network principles to neural architecture:
emergence of spatial-opponent cells. Proc Natl Acad Sci USA.
1986;83:7508–12.
3. Linsker R. From basic network principles to neural architecture:
emergence of orientation-selective cells. Proc Natl Acad Sci USA.
1986;83:8390–4.
4. Linsker R. From basic network principles to neural architecture: emergence of orientation columns. Proc Natl Acad Sci USA. 1986;83:8779–83.
5. Domany E, Hemmen JL van, Schulten K: Models of neural networks III.
Berlin: Springer; 2012.
6. Schall JD, Vitek DJ, Leventhal AG. Retinal constraints on orientation specificity in cat visual cortex. J Neurosci. 1986;6(3):823–36.
7. Wang Y, Brzozowska-Prechtl A, Karten HJ. Laminar and columnar auditory
cortex in avian brain. Proc Natl Acad Sci USA. 2010;107:12676–81.
P192
How do hidden units shape effective connections
between neurons?
Braden A. W. Brinkman1,2, Tyler Kekona1, Fred Rieke2,3, Eric Shea‑Brown1,2,4,
Michael Buice4
1
Department of Applied Mathematics, University of Washington, Seattle,
WA 98195, USA; 2Department of Physiology and Biophysics, University
of Washington, Seattle, WA 98195, USA; 3Howard Hughes Medical
Institute, University of Washington, Seattle, WA 98195, USA; 4Allen
Institute for Brain Science, Seattle, WA, 98109, USA
Correspondence: Braden A. W. Brinkman ‑ bradenb@uw.edu
BMC Neuroscience 2016, 17(Suppl 1):P192
A major challenge in neuroscience is understanding how “hidden
units”—neurons or other influences not observed in an experiment—
influence the behavior of the observed neurons. Much work has
been done on the inferring network interactions from data [1, 2], but
it remains unknown how hidden neurons shape the network interactions inferred. Using techniques from non-equilibrium statistical
physics, we have developed a theoretical framework to predict how
effective connections in subsampled networks depend on the true
connections in the full network. Beyond calculating effective connections, this approach can be systematically expanded to study how hidden units generate effective noise in subsampled networks.
As an example, we apply this framework to a network of three spiking neurons described by a generalized linear model (GLM) with rates
driven by a neuron’s own filtered spiking activity and those from which
it receives input. By approximating the subsampled network as a GLM
with effective spike-filters corrupted by Gaussian noise, we can analytically calculate how hidden units transform the filters (Fig. 100) and
give rise to correlations in the effective noise (not shown). Based on
our 3-neuron results, we conjecture that for general networks within
this framework the filter between neurons i and j is modified by
Page 108 of 112
Fig. 100 A Self-history filters (diagonal) and directed coupling filters
between neurons (off-diagonal) in the full 3-neuron network. Neuron
1 is excitatory and its couplings to the other neurons are strictly
positive. Neurons 2 and 3 are inhibitory and make strictly negative
couplings to other neurons. There is no coupling from neuron 2–3. B
Effective self-history filters (diagonals) and coupling filters (off-diagonals) when neuron 3 is hidden. The bottom row is unaltered because
neuron 2 makes no coupling to neuron 3. The filters in the top row
are changed due to the influence of signals neuron 1 sends to itself
through neuron 3 and to neuron 2 through neuron 3. Although neuron 1’s true self-history filter and coupling from neuron 2 are negative
the effective filters change sign. C The effective self-history filter of
neuron 1 when both neurons 2 and 3 are hidden. Times and filter
amplitudes are given in arbitrary units (a.u.)
corrections from every path that neuron i can send a signal to neuron j
through hidden units.
Acknowledgements: Support provided by the Sackler Scholar Program in Integrative Biophysics (BAWB), CRCNS grant DMS-1208027
(ESB, FR), NSF-DMS-1056125 (ESB), NIH grant EY11850 (FR), HHMI (FR).
ESB and MB thank the Allen Institute founders, Paul G. Allen and Jody
Allen, for their vision, encouragement and support.
References
1. Pillow JP, Shlens J, Paninski L, Sher A, Litke AM, Chichilnisky EJ, Simoncelli
EP. Spatio-temporal correlations and visual signalling in a complete
neuronal population. Nature. 2008;454:995–9.
2. Pillow JW, Latham P. Neural characterization in partially observed populations of spiking neurons. Adv Neural Inf Process Syst. 2007;3256:1–8.
P193
Characterization of neural firing in the presence
of astrocyte‑synapse signaling
Maurizio De Pittà1,2, Hugues Berry2,3, Nicolas Brunel1,3
1
Department of Neurobiology, University of Chicago, Chicago, IL 60637,
USA; 2Project‑Team BEAGLE, INRIA Rhône‑Alpes, Villeurbanne, F‑69603,
France; 3Department of Statistics, University of Chicago, Chicago, IL
60637, USA
Correspondence: Maurizio De Pittà ‑ maurizio.depitta@gmail.com
BMC Neuroscience 2016, 17(Suppl 1):P194
We study analytically the dynamics of neural activity in the presence
of synaptic inputs modulated by astrocyte-released neurotransmitters (i.e. the so-called “gliotransmitters”). We start with the simple
scenario of gliotransmitter-mediated modulation of synaptic release
probability at N excitatory synapses impinging on a single postsynaptic neuron as well as on the same astrocyte domain. In this scenario,
release from pre-synaptic terminals leads to activation of the astrocyte, that in turn modulates synaptic release through gliotransmitter
release. In the limit of N → ∞ synapses, we derive equations relating
gliotransmitter release to the instantaneous presynaptic rate, identify
conditions for co-existence of multiple states of synaptic release, and
study their stability. In the bistable regime, long-lasting potentiation
of synaptic release by gliotransmission accounts for the emergence of
persistent postsynaptic firing. Analysis of the coefficient of variation
(CV) of the ensuing interspike interval distribution reveals increased
BMC Neurosci 2016, 17(Suppl 1):54
firing variability following stimulation and in the presence of gliotransmission, in close analogy with increased CV values experimentally
observed during the delay period in working-memory related tasks.
We then extend our analysis to the scenario of a balanced neural network coupled with a network of astrocytes, and demonstrate the existence of an analogous mechanism for persistent neural firing by mean
field theory. Taken together, our analysis suggests a novel astrocytebased mechanism for persistent activity, and provides experimentally
testable hypotheses on the possible involvement of astrocytes in cognitive tasks related to working memory.
P194
Metastability of spatiotemporal patterns in a large‑scale network
model of brain dynamics
James A. Roberts1, Leonardo L Gollo1, Michael Breakspear1
1
Systems Neuroscience Group, QIMR Berghofer Medical Research
Institute, Brisbane, QLD 4006, Australia
Correspondence: James A. Roberts ‑ james.roberts@qimrberghofer.edu.
au
BMC Neuroscience 2016, 17(Suppl 1):P194
Advances in mapping the human connectome have yielded increasingly-detailed descriptions of large-scale brain networks, prompting
growing interest in the dynamics that emerge from this structural
connectivity. Moreover, there is a desire to move beyond simple
static functional connectivity measures to better describe and understand the more complex repertoire of brain dynamics, which unfolds
on multiple time scales. Here, we analyze the dynamics that emerge
from a neural mass model [1, 2] with network connectivity derived
from densely-seeded probabilistic tractography on human diffusion imaging data [3]. We find a rich array of three-dimensional wave
patterns, including traveling waves, spiral waves, sources, and sinks
(Fig. 101). These patterns are metastable, with the dynamics cycling
between several relatively long-lived states. Varying the overall coupling strength and coupling delay reveals a complex parameter space,
with other emergent patterns such as cycling between strongly-correlated clusters and multistability between different regimes (as distinct
from metastability within a single regime). These dynamics accord
with empirical data from multiple imaging modalities, including
Fig. 101 Large-scale wave patterns for strong coupling, showing
four time snapshots for a traveling wave (top), a spiral wave (middle),
and a sink pattern (bottom). Warmer colors denote higher amplitudes
Page 109 of 112
observations of electrical waves in cortical tissue [4] and the presence
of sequential spatiotemporal patterns in resting state MEG data [5].
By characterizing the dynamic states and time scales in our simulated
data, we demonstrate the richness of dynamics that emerge from the
human connectome. This work lays a platform for detailed analyses
of large-scale functional neuroimaging data and their mechanistic
underpinnings.
References
1. Breakspear M, Terry JR, Friston KJ: Modulation of excitatory synaptic coupling facilitates synchronization and complex dynamics in a biophysical
model of neuronal dynamics. Network. 2003;14:703–32.
2. Gollo LL, Zalesky A, Hutchison RM, van den Heuvel M, Breakspear M.
Dwelling quietly in the rich club: Brain network determinants of slow cortical fluctuations. Philos Trans R Soc Lond B Biol Sci. 2015;370:20140165.
3. Roberts JA, Perry A, Lord AR, Roberts G, Mitchell PB, Smith RE, Calamante
F, Breakspear M. The contribution of geometry to the human connectome. NeuroImage. 2016;124:379–93.
4. Townsend RG, Solomon SS, Chen SC, Pietersen AN, Martin PR, Solomon
SG, Gong P. Emergence of complex wave patterns in primate cerebral
cortex. J Neurosci. 2015;35:4657–62.
5. Baker AP, Brookes MJ, Rezek IA, Smith SM, Behrens T, Smith PJ, Woolrich
M. Fast transient networks in spontaneous human brain activity. eLife.
2014;3:e01867.
P195
Comparison of three methods to quantify detection
and discrimination capacity estimated from neural population
recordings
Gary Marsat1, Jordan Drew1, Phillip D. Chapman1, Kevin C. Daly1, Samual
P. Bradley1
1
Department of Biology, West Virginia University, Morgantown, WV 26506,
USA
Correspondence: Gary Marsat ‑ gary.marsat@mail.wvu.edu
BMC Neuroscience 2016, 17(Suppl 1):P195
The responses of sensory neurons carry information about the presence and the identity of relevant external events [1]. The pattern of
activity in populations of such neurons must be decoded by post
synaptic networks for the information to contribute to the elaboration of behavioral responses. For two stimuli to be discriminated, or a
stimulus discriminated from background (i.e. detected), the patterns
of activity it elicits in the encoding neural population must be different enough that target decoders are activated differentially [2]. In this
research we used recording from electrosensory neurons in Gymnotid
fish [3] and recordings from the antennal lobe of moth [4] to compare
three methods that can be used to quantify how accurately the neural responses can support detection and discrimination tasks. The first
two methods, namely Euclidian distances [5] and spike metrics distances [3], have traditionally been used by neurophysiologist to characterize the information carried by spike trains and to compare the
responses to different stimuli. The third method we explored relies on
the clustering neural network provided in Matlab toolboxes that uses
unsupervised learning. Neural networks of this type are widely used
by engineers to perform practical tasks but are rarely used by neuroscientist to study actual neural systems. The clustering tool relies on
a quantification similar in many ways to Euclidian distances or spike
distance metrics. However, since it learns to weight the inputs to
allow optimal clustering, the weight patterns of networks that cluster accurately can reveal features that actual neural networks ought
to have (including synaptic facilitation and depression or amount of
convergence of inputs). We show that, both for the olfactory system
of moth and the electrosensory system of fish, the neural network
decoder outperforms the other two decoding analyses by weighting
more heavily information rich inputs and more weakly noisy ones. We
argue that this tool can be advantageously used to quantify neural
coding, can make testable prediction regarding the characteristics
of the decoding network, and, most importantly, can be easily used
and implemented by researchers who have little training in neural
modeling.
BMC Neurosci 2016, 17(Suppl 1):54
Page 110 of 112
Acknowledgements: This work was supported by NSF Grant IOS1557846 to G.M.
References
1. Ollerenshaw DR, Zheng HJV, Millard DC, Wang Q, Stanley GB. The adaptive trade-off between detection and discrimination in cortical representations and behavior. Neuron. 2014;81(5):1152–64.
2. Clemens J, Ronacher B. Feature extraction and integration underlying perceptual decision making during courtship behavior. J Neurosci.
2013;33(29):12136–45.
3. Marsat G, Maler L. Neural heterogeneity and efficient population codes
for communication signals. J Neurophysiol. 2010;104(5):2543–55.
4. Staudacher EM, Huetteroth W, Schachtner J, Daly KC. A 4-dimensional
representation of antennal lobe output based on an ensemble of characterized projection neurons. J Neurosci Methods. 2009;180(2):208–23.
5. Daly KC, Bradley S, Chapman PD, Staudacher EM, Tiede R, Schachtner J.
Space takes time: concentration dependent output codes from primary
olfactory networks rapidly provide additional information at defined
discrimination thresholds. Front Cell Neurosci. 2016;9.
P196
Quantifying the constraints for independent evoked
and spontaneous NMDA receptor mediated synaptic transmission
at individual synapses
Sat Byul Seo1, Jianzhong Su1, Ege T. Kavalali2, Justin Blackwell1
1
Department of Mathematics, University of Texas at Arlington, Arlington,
TX 76019, USA; 2Department of Neuroscience, University of Texas
Southwestern Medical Center, Dallas, TX 75390, USA
Correspondence: Sat Byul Seo ‑ satbyul.seo@mavs.uta.edu
BMC Neuroscience 2016, 17(Suppl 1):P196
Presynaptic terminals release neurotransmitters either in response
to action potentials or spontaneously independent of presynaptic
activity. In the case of glutamate, released neurotransmitters activate
N-methyl-D-asparate (NMDA) receptors within a single postsynaptic site and give rise to miniature postsynaptic currents. In this study,
we used a mathematical model to simulate spontaneous and evoked
neurotransmission processes resulting from glutamate release within
a synapse and evaluate the quantitative constraints that determine
their degree of overlap independent signaling mediated by spontaneous and evoked release events. First we simulated isotropic diffusion of 4000 glutamates molecules release from a point source. We
then simulated release of the glutamate molecules through a vesicle
by addition of two compartments that one modeled the vesicle and
the other represented the fusion pore. After we obtains the glutamate
concentration from the standard heat equation then determine the
opening probability of individual receptor using a state model (3C2O).
Those two problems in MATLAB are solved.
If we assume a fivefold–tenfold ratio as a good indicator for independent currents, then we cannot assure independency with the
structure for medium and small synapses in our current hypothesis.
Figure 102 shows that small synapse (200 nm × 200 nm) might not
have independent signaling when glutamate release instantaneously
because evoked and spontaneous receptors are not far away from
each other and thus not far from the release site in either evoked or
spontaneous releases, the ratio of open probability is close to 1. The
open probability is consistent up to 90 nm far from the release site as
in Fig. 102. However for small synapses, as glutamate release through
10 and 2 nm vesicle fusion pore, the open probability ratio decreases
more drastically and become close to zero, and in 2 nm pore, the ratio
achieves 10-fold reduction at 90 nm distance, giving plausibility for
independent signaling.
Conclusion From the results we conclude that peak open value is
most sensitive to the distance from the receptor to the release site.
Glutamate release speed or fusion pore size is relevant but to a lesser
degree. The calculation was first performed for a large synapse of
0.36 µm2 (with R6 near the center for evoked neurotransmission, and
R16 for spontaneous neurotransmission) which established in theory
that two non-overlapping domains that give rise to independent signaling in large synapses [1]. Then calculations in medium size 0.16 µm2
Fig. 102 In small synapses (200 nm × 200 nm), Ratios of maximum
NMDA receptor opening probabilities as functions of receptor
distance for different release speed (slow, 2 nm fusion pore—triangle,
regular, 10 nm fusion pore—asterisk, and instantaneous—circle) of
glutamate vesicle release. The open probabilities were calculated by
the kinetics equation, when glutamates are released above the center
location
and small size 0.04 µm2 push the biophysical envelope for independent currents, as the degree of independence decreases when the size
of synapse gets smaller, or the distances from evoked and spontaneous receptors to the release site are closer together.
Reference
1. Atasoy D, Ertunc M, Moulder KL, et al. Spontaneous and evoked glutamate release activates two populations of NMDA receptors with limited
overlap. J Neurosci. 2008;28:10151–166.
P199
Gamma oscillation via adaptive exponential integrate‑and‑fire
neurons
LieJune Shiau1, Laure Buhry2, Kanishka Basnayake3
1
Department of Mathematics, University of Houston, Clear Lake, Houston,
TX, 77059, USA; 2Department of Computational Neurosciences, University
of Lorraine, Nancy, 54600, France; 3Computational Neurosciences
Laboratory, Ecole Polytechnique Federale de Lausanne, CH‑1015,
Switzerland
Correspondence: LieJune Shiau ‑ shiau@uhcl.edu
BMC Neuroscience 2016, 17(Suppl 1):P199
Coherent oscillation of neuronal spiking in the brain is known related
to cognitive functions, including perception, attention, and memory.
It is therefore important to determine the properties of neurons and
network architectures in emerging the coherent activities that influence the network collective behaviors. It is known that, in local cortical
circuits, the probability for any pair of pyramidal cells to be connected
is low and about 0.1–0.2 [1]. Wang and Buzsaki [2] numerically demonstrate that, in a heterogeneous and inhibitory network with sparse
and random connection, the minimal connection required per neuron
to observe coherence oscillation is approximately 60 with Hodgkin–
Huxley (H–H) type interneurons in certain parameter regime. More
importantly, this minimal number is relatively independent to the network size. In contrast, Golomb and Hanel [3] theoretically show that,
BMC Neurosci 2016, 17(Suppl 1):54
identical inhibitory neurons in a sparse and random network, the minimal connection required per neuron to exhibit coherence oscillation
is about 360 with integrate-and-fire (IF) neurons. The minimal connection required in either study depends on the intrinsic and synaptic
properties of the neurons.
It is shown that hippocampal CA1 neurons make an average of about
60 contacts to other neurons within a spatial span of approximately
500 μm [4]. Hence, to study rhythmic oscillation in hippocampal
networks, H–H type neurons, instead of IF neurons, could produce
Gamma rhythm through sparsely connected network in a population
of interneurons. These findings are believed to address the importance
of detailed physiological properties of single neurons in determining
collective network behaviors.
We adopt an increasingly popular two-dimensional adaptive exponential integrate-and-fire (aEIF) model [5] which is equipped with a
subthreshold adaptation coupling the voltage and a slow current, and
a spike-triggered adaptation regulated through each spike. To demonstrate that aEIF neurons can provide adequate networks in inducing
Gamma frequency as in hippocampus effectively, we establish the
minimal synaptic contacts required in sparse and random networks of
aEIF neurons to exhibit coherent oscillations, and the impacts of neuronal and synaptic properties have on the minimal value. The aEIF neuron provides more physiological neuronal details than IF neuron, but
much less than the H–H neurons. Intuitively, it may be anticipated that
the minimal synaptic contacts of aEIF required in such networks lies
somewhere between that of the IF and H–H neurons.
We demonstrate that the minimal synaptic contacts required in such
networks of aEIF neurons to exhibit Gamma rhythm is also surprisingly
low with an approximation of 60 in certain parameter regime. More
specifically, the minimal connection required per neuron for the onset
of network synchrony is not a faction of the total number of network
neurons. It either remains constant or only depends weakly on the
total network neurons. This study indicates that the inclusion of subthreshold and spike-triggered adaptations provides aEIF neuron with
features to compensate for the lack of physiological details, as supposed to its H–H neuron counterpart, in studying Gamma rhythm in
the brain. Our result is very encouraging in building neural network
studies through a simple two-dimensional model.
References
1. Song S, Sjostrom P, Reigl M, Nelson S, Chklovskii D. Highly nonrandom
features of synaptic connectivity in local cortical circuits. PLoS Biol.
2005;3(3):0507–19.
2. Wang X, Buzsaki G. Gamma oscillation by synaptic inhibition in
a hippocampal interneuronal network model. J Neurophysiol.
1996;20:6402–13.
3. Golomb D, Hansel D. The number of synaptic inputs and the synchrony of
large sparse neuronal networks. Neural Comput. 1999;12(5):1095–1139.
4. Sik A, Penttonen M, Ylinen A, Buzsaki G. Hippocampal CA1 interneurons:
an in vivo intracellular labeling study. J Neurosci. 1999;24:49–65.
5. Brette R, Wulfram G. Adaptive exponential integrate-and-fire model
as an effective description of neuronal activity. J Neurophysiol.
2005;94:3637–42.
P200
Visual face representations during memory retrieval compared
to perception
Sue‑Hyun Lee1,2, Brandon A. Levy3, Chris I. Baker3
1
Department of Bio and Brain Engineering, Korea Advanced Institute
of Science and Technology (KAIST), Daejeon 34141, Republic of Korea; 2
Program of Brain and Cognitive Engineering, Korea Advanced Institute
of Science and Technology (KAIST), Daejeon 34141, Republic of Korea;
3
Laboratory of Brain and Cognition, National Institute of Mental Health,
National Institutes of Health, Bethesda, MD 20892, USA
Correspondence: Sue‑Hyun Lee ‑ suelee@kaist.ac.kr
BMC Neuroscience 2016, 17(Suppl 1):P200
In our daily life, we can easily discriminate and recognize familiar
faces. Much evidence suggests that the fusiform face area (FFA) and
the occipital face area (OFA) are involved in face processing [1–3].
Page 111 of 112
However, it remains unclear how individual face information is represented in the visual cortex during retrieval compared to perception. To
address this question, we performed an event-related functional magnetic resonance imaging (fMRI) experiment, comprising separate perception, learning and retrieval sessions. During the perception session,
which took place inside the scanner, participants were presented with
fixed pairings of six auditory cues (pseudowords) with six face images,
and six auditory cues with six shoe images. During the learning session, which took place on a separate day outside the scanner, participants were trained to memorize the pseudoword-image associations
for about 1 h. Finally, 1 day after the learning session, participants were
scanned again and instructed to retrieve each image in response to
auditory presentation of the paired pseudoword cue. To test the veracity of the retrieved visual information, participants were asked to perform forced-choice tests after the retrieval scan session, in which they
heard one of the pseudoword cues and chose the paired category
or image. Every participant showed near perfect performance in the
forced-choice test. We focused on the patterns of response in faceselective cortical areas. Using multivoxel pattern analyses, we found
that FFA showed more discriminable patterns of response to individual faces during retrieval compared to those elicited during perception. In contrast object-selective areas, which respond well to images
of shoes, did not show any significant difference between perception
and retrieval for individual shoe images. To determine whether the
increased discrimination reflected a difference between perceived and
retrieved face information and not an effect of learning, we conducted
a similar fMRI experiment in which the second session was also perception and not retrieval. Importantly, there was no difference in face
discrimination between the first and second perception sessions in
FFA. Taken together, these results suggest that retrieval of face information generates more discriminative neural responses for individual
faces than that evoked by perception of the very same faces.
Acknowledgements: This work was supported by the US National
Institutes of Health Intramural Research Program of the National Institute of Mental Health, and a NARSAD Young Investigator Grant from
the Brain & Behavior Research Foundation.
References
1. Kanwisher N, McDermott J, Chun MM. The fusiform face area: a module
in human extrastriate cortex specialized for face perception. J Neurosci.
1997;17:4302–11.
2. Tarr MJ, Gauthier I. FFA: a flexible fusiform area for subordinate-level visual
processing automatized by expertise. Nat Neurosci. 2000;3:764–9.
3. Kanwisher N, Yovel G. The fusiform face area: a cortical region specialized for the perception of faces. Philos Trans R Soc Lond B Biol Sci.
2006;361:2109–28.
P201
Top‑down modulation of sequential activity within packets
modeled using avalanche dynamics
Timothée Leleu1, Kazuyuki Aihara1
1
Institute of Industrial Science, the University of Tokyo, Tokyo, Japan
Correspondence: Timothée Leleu ‑ timothee@sat.t.u‑tokyo.ac.jp
BMC Neuroscience 2016, 17(Suppl 1):P201
Recent experiments show that short activity packets are triggered
by external stimuli or internal spontaneous events during which the
temporal order of spikes is only partially stereotypical [1]. Moreover, it
has been suggested that the timing of neurons during these packets
depends on top-down modulatory inputs that “gate” the sensory information and represents either the replay of previously stored patterns
or information about ongoing external stimuli [1]. Finally, it has been
observed that spontaneous activity consists in the superposition of
multiple overlapping packets [1]. We propose a simple model of cortical neural networks that reproduces these experimental observations
[1] and an analytical description of the top-down modulation of packets using avalanche dynamics [2]. The proposed theory allows predicting the average size of packets using the synaptic weight matrix and
vice versa.
BMC Neurosci 2016, 17(Suppl 1):54
Page 112 of 112
Fig. 104 A Auto-encoder NN coupled to the VN. B, C Depict desynchronized (ɛ = 1) and synchronized (ɛ = 0) states of VN respectively.
The corresponding output weight patterns learnt by the autoencoder, driven by the VN, trained on bar pattern data (B1, C1) and
MNIST data (B2, C2)
Fig. 103 A1–A4 Network structure. Neurons of the first and second
stored pattern are represented by colors ranging from blue to green
and yellow to red, respectively. Effective synaptic connections can be
calculated and are shown by colored segments. B1–B4 Cross-correlograms (CCG) of single neuron activity with the summed activity of
other neurons (see [1]). C1–C4 Center of mass of CCGs, noted μCC
The model describes the neural activity within a cortical area that
receives top-down modulatory and bottom-up sensory inputs from
higher-order areas and thalamic projections, respectively. The activity of excitatory neurons is simulated using the leaky integrate-andfire model. Excitatory synaptic connections are modified on shorter
and longer time-scales by short-term depression and spike-timing
dependent plasticity, respectively. Sequential patterns are stored
within the recurrent connections of the middle area by repeated presentation of the external stimuli.
Figure 103A1–C1, A2–C2 show that only the first and second stored
sequences are replayed when the first and second top-down input is
active, respectively, although the bottom-up inputs trigger the starting neurons of both sequential patterns. When there is no top-down
input, the spontaneous activity is composed of the time-compressed
superposition of both sequential patterns (see Fig. 103A4–C4).
Acknowledgements: This research was supported by ImPACT Program of Council for Science, Technology and Innovation (Cabinet
Office, Government of Japan).
References
1. Luczak A, Bartho P, Harris KD. Gating of sensory input by spontaneous
cortical activity. J Neurosci. 2013;33(4):1684–95.
2. Leleu T, Aihara K. Unambiguous reconstruction of network structure
using avalanche dynamics. Phys Rev E. 2015;91:022804.
Q28
An auto‑encoder network realizes sparse features under the
influence of desynchronized vascular dynamics
Ryan T. Philips1, Karishma Chhabria1, V.Srinivasa Chakravarthy1
1
Department of Biotechnology, Indian Institute of Technology, Madras,
Chennai 600036, India
Correspondence: V. Srinivasa Chakravarthy ‑ schakra@iitm.ac.in
BMC Neuroscience 2016, 17(Suppl 1):Q28
Please note that this abstract was presented at the previous year’s
24th Annual Computational Neuroscience Meeting: CNS-2015.
Cerebral vascular dynamics are generally thought to be controlled
by neural activity in a unidirectional fashion. However, both computational modeling and experimental evidence points to the feedback
effects of vascular activity on neural dynamics [1, 2]. Vascular feedback
in the form of glucose and oxygen controls neuronal ATP, which in turn
can control the threshold of neural firing. We present a computational
model of a neuro-vascular system in which a network of ‘vascular units’
supply ‘energy’ to a neural network (NN), which reduces neural firing
threshold. The vascular network (VN) is modeled by a network of oscillators as in [3]. Neuronal pools fed by the complex dynamics of VN are
turned ON and OFF randomly. We show that such a feedback mechanism results in sparse weight matrix, thereby enhancing the performance of an auto-encoder NN.
In the proposed NN model, the hidden layer is coupled to a vascular
network in a one-to-one fashion (Fig. 104A) and is trained using backpropagation. The cross-entropy (ce) error measure is used to update
the energy demand parameter (Md) which is fed back to the VN. Md in
turn governs the state of the vascular units, which determines if the
neuron should be turned ON/OFF. This paradigm of randomly turning neural units ON/OFF is adapted from [4]. High Md results in an
increase in ce as the neuronal dropout level is too low; similarly low
Md results in an increase in ce due to high dropout. The time scale of
vascular dynamics (Md) is much longer than that of neural dynamics
(ce), reflecting physiology. The network was trained on two datasets:
overlapping bar patterns and MNIST data. The settled weight matrices corresponding to both the synchronized (Fig. 104B) and desynchronized vascular dynamics (Fig. 104C) are shown in the Fig. 104B1,
B2, C1, C2, respectively. Our earlier modeling study highlighted the
link between desynchronized vascular dynamics and efficient energy
delivery in skeletal muscle [3]. We now show that desynchronized vascular dynamics leads to efficient training in an auto-encoder NN.
References
1. Chander BS, Chakravarthy VS. A computational model of neuro-gliovascular loop interactions. PloS One. 2012;7(11):e48802.
2. Moore CI, Cao R: The hemo-neural hypothesis: on the role of blood flow
in information processing. J Neurophysiol. 2008;99(5):2035.
3. Pradhan RK, Chakravarthy V, Prabhakar A. Effect of chaotic vasomotion in
skeletal muscle on tissue oxygenation. Microvasc Res. 2007;74(1):51–64.
4. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. J Mach
Learn Res. 2014;15(1):1929–58.