Available online at www.sciencedirect.

com

ScienceDirect

Towards neural co-processors for the brain: combining

decoding and encoding in brain–computer interfaces
Rajesh PN Rao

The field of brain–computer interfaces is poised to advance After a promising start, there was a surprising lull in the
from the traditional goal of controlling prosthetic devices using field until the 1990s when, spurred by the advent of multi-
brain signals to combining neural decoding and encoding electrode recordings as well as fast and cheap computers,
within a single neuroprosthetic device. Such a device acts as a the field saw a resurgence under the banner of brain–
‘co-processor’ for the brain, with applications ranging from computer interfaces (BCIs; also known as brain–machine
inducing Hebbian plasticity for rehabilitation after brain injury to interfaces and neural interfaces) [1,2].
reanimating paralyzed limbs and enhancing memory. We
review recent progress in simultaneous decoding and encoding A major factor in the rise of BCIs has been the application
for closed-loop control and plasticity induction. To address the of increasingly sophisticated machine learning techni-
challenge of multi-channel decoding and encoding, we ques for decoding neural activity for controlling prosthetic
introduce a unifying framework for developing brain co- arms [8,9,10
], cursors [11,12,13
,14,15,16

], spellers
processors based on artificial neural networks and deep [17,18] and robots [19–22]. Simultaneously, researchers
learning. These ‘neural co-processors’ can be used to jointly have explored how information can be biomimetically or
optimize cost functions with the nervous system to achieve artificially encoded and delivered via stimulation to neu-
desired behaviors ranging from targeted neuro-rehabilitation to ronal networks in the brain and other regions of the
augmentation of brain function. nervous system for auditory [23], visual [24], propriocep-
tive [25], and tactile [26,27,28
,29,30] perception.
Address
Center for Neurotechnology, Paul G. Allen School of Computer Science Building on these advances in neural decoding and
and Engineering, University of Washington, Seattle, United States
encoding, researchers have begun to explore bi-
Corresponding author: Rao, Rajesh PN (rao@cs.washington.edu) directional BCIs (BBCIs) which integrate decoding
and encoding in a single system. In this article, we
review how BBCIs can be used for closed-loop control
Current Opinion in Neurobiology 2019, 55:142–151
of prosthetic devices, reanimation of paralyzed limbs,
This review comes from a themed issue on Machine learning, big restoration of sensorimotor and cognitive function,
data, and neuroscience
neuro-rehabilitation, enhancement of memory, and brain
Edited by Maneesh Sahani and Jonathan Pillow augmentation.

Motivated by this recent progress, we propose a new

https://doi.org/10.1016/j.conb.2019.03.008
unifying framework for combining decoding and encod-
ing based on ‘neural co-processors’ which rely on artifi-
0959-4388/ã 2019 Elsevier Ltd. All rights reserved.
cial neural networks and deep learning. We show that
these ‘neural co-processors’ can be used to jointly
optimize cost functions with the nervous system to
achieve goals such as targeted rehabilitation and aug-
mentation of brain function, besides providing a new
tool for testing computational models and understand-
Introduction
ing brain function [31].
A brain–computer interface (BCI) [1–4] is a device that
can (a) allow signals from the brain to be used to control
devices such as prosthetics, cursors or robots, and (b) allow
external signals to be delivered to the brain through Simultaneous decoding and encoding in
neural stimulation. The field of BCIs has made enormous BBCIs
strides in the past two decades. The genesis of the field Closed-loop prosthetic control
can be traced to early efforts in the 1960s by neuroscien- Consider the problem of controlling a prosthetic hand
tists such as Fetz [5
] who studied operant conditioning in using brain signals. This involves (1) using recorded
monkeys by training them to control the movement of a neural responses to control the hand, (2) stimulating
needle in an analog meter by modulating the firing rate of somatosensory neurons to provide tactile and propriocep-
a neuron in their motor cortex. Others such as Delgado tive feedback, and (3) ensuring that stimulation artifacts
and Vidal explored techniques for neural decoding and do not corrupt the recorded signals being used to control
stimulation in early versions of neural interfaces [6
,7]. the hand. Several artifact reduction methods have been

Current Opinion in Neurobiology 2019, 55:142–151 www.sciencedirect.com

 Toward neural co-processors for the brain Rao 143

proposed for (3) – we refer the reader to Refs. [32–34]. We find a target object that delivers stimulation similar to a
focus here on combining (1) decoding with (2) encoding. control object. In their experiment, the monkey con-
trolled a virtual arm using intracortical recordings from
Most state-of-the-art decoding algorithms for intracortical posterior parietal cortex and a Kalman-filter-based decod-
BCIs are based on a linear decoder such as the Kalman ing scheme where the Kalman filter’s state was defined as
filter. Typically, the state vector x for the Kalman filter is the virtual hand’s position, velocity and acceleration in
chosen to be a vector of kinematic quantities to be three dimensions. The encoding algorithm involved stim-
estimated, such as hand position, velocity, and accelera- ulating S1 via three closely located electrodes using a
tion. The likelihood (or measurement) model for the 300 Hz biphasic pulse train for up to one second while the
Kalman filter specifies how the kinematic vector xt at virtual hand held the object. After training, the monkey
time t relates linearly (via a matrix B) to the measured was able to move the virtual hand to the correct target
neural activity vector yt: with success rates ranging from approximately 70% to
more than 90% over the course of eight days (chance level
was 50%).
y t ¼ Bx t þ mt
Finally, Flesher et al. [38

] have recently shown that a
while a dynamics model specifies how xt linearly changes paralyzed patient can use a bidirectional BCI for closed-loop
(via matrix A) over time: control of a prosthetic hand in a continuous force matching
task. Control signals were decoded from multi-electrode
recordings in M1 using a linear decoder that mapped M1
x t ¼ Ax t1 þ nt
firing rates to movement velocities of the robotic arm. Initial
training data for the linear decoder was obtained by asking the
nt and mt are zero-mean Gaussian noise processes. The subject to observe the robotic hand performing hand shaping
Kalman filter computes the optimal estimates for kine- tasks such as ‘pinch’ (thumb/index/middle flexion–exten-
matics xt (both mean and covariance) given current and all sion), ‘scoop’ (ring/pinky flexion/extension) or grasp (all finger
past neural measurements. flexion) and recording M1 firing rates, followed by a second
training phase involving computer-assisted control to fine
One of the first studies to combine decoding and encod- tune the decoder weights. The subject then performed a 2D
ing was by O’Doherty et al. [35] who showed that stimu- force matching task with the robotic hand using the trained
lation of somatosensory cortex could be used to instruct a decoder to pinch, scoop or grasp a foam object either gently or
rhesus monkey which of two targets to move a cursor to; firmly while using stimulation of S1 to get feedback on the
the cursor was subsequently controlled using a BCI based force applied. The encoding algorithm linearly mapped
on linear decoding to predict the X-coordinate and torque sensor data from the robotic hand’s finger motors to
Y-coordinate of the cursor. A later study by the same pulse train amplitude of those stimulating electrodes that
group [36

] demonstrated true closed-loop control. Mon- previously elicited percepts on the corresponding fingers of
keys used a BCI based on primary motor cortex (M1) the subject. The researchers showed that the subject was able
recordings and Kalman-filter-based decoding to actively to continuously control the flexion/extension of the pinch and
explore virtual objects on a screen with artificial tactile scoop dimensions while evaluating the applied torque based
properties. The monkeys were rewarded if they found the on force feedback from S1 stimulation. The success rate for
object with particular artificial tactile properties. During pinch, scoop, or grasp with gentle or firm forces was signifi-
brain-controlled exploration of an object, the associated cantly higher with stimulation feedback compared to feed-
tactile information was delivered to somatosensory cortex back from vision alone.
(S1) via intracortical stimulation. Tactile information was
encoded as a high-frequency biphasic pulse train (200 Hz Reanimating paralyzed limbs
for rewarded object, 400 Hz for others) presented in Rather than controlling a prosthetic limb, BBCIs can also be
packets at a lower frequency (10 Hz for rewarded, 5 Hz used to control electrical stimulation of muscles to restore
for unrewarded objects). Because stimulation artifacts movement in a paralyzed limb. Moritz et al. [39

] demon-
masked neural activity for 5–10 ms after each pulse, an strated this approach in two monkeys by translating the
interleaved scheme of alternating 50 ms recording and activity of single motor cortical neurons into electrical stimu-
50 ms stimulation was used. The monkeys were able to lation of wrist muscles to move a cursor on a computer screen.
select the desired target object within a second or less The decoding scheme involved operant conditioning to
based only on its tactile properties as conveyed through volitionally control activity of a motor cortical neuron to
stimulation. initially move a cursor into a target. After training, the activity
from the motor cortical neuron was converted into electrical
Klaes et al. [37] have also demonstrated that a monkey can stimuli which was delivered to the monkey’s temporarily
utilize intracortical stimulation in S1 to perform a match- paralyzed wrist muscles (this type of stimulation is called
to-sample task where the goal is to move a virtual arm and functional electrical stimulation, or FES). Flexor FES current

www.sciencedirect.com Current Opinion in Neurobiology 2019, 55:142–151

144 Machine learning, big data, and neuroscience

was set to be proportional to the rate above a threshold designed an implantable BBCI called the stimoceiver that
(0.8  [firing rate  24] with a maximum of 10 mA), and could communicate with a computer via radio. Delgado
extensor FES was inversely proportional to the rate below was the first to combine decoding with encoding to shape
a second threshold (0.6  [12  firing rate] with a maximum of behavior: his decoding algorithm detected spindles in the
10 mA). Both monkeys were able to modulate the activity of amygdala of a monkey and for each detection, triggered
cortical neurons to control their paralyzed wrist muscles and stimulation in the reticular formation, which is associated
move a manipulandum to acquire five targets. Ethier et al. [40] with negative reinforcement. After six days, spindle
extended these results to grasping and moving objects using a activity was reduced to one percent of normal levels,
linear decoder with a static nonlinearity applied to about making the monkey quiet and withdrawn. Unfortunately,
100 neural signals from M1. efforts to extend this approach to humans to treat depres-
sion and other disorders yielded inconsistent results.
Extending the approach to humans, Bouton et al. [41

]
showed that a quadriplegic man with a 96-electrode array Delgado’s work did eventually inspire commercial brain
implanted in the hand area of the motor cortex could use implants such as Neuropace’s RNS system that detects
cortical signals to electrically stimulate muscles in his onset of seizures using time-based and frequency-based
paralyzed forearm and produce six different wrist and hand methods from brain surface recordings (ECoG) and sti-
motions. For decoding, six separate support vector mulates the region where the seizure originates. Also
machines were applied to mean wavelet power features inspired by Delgado’s work is the technique of deep
extracted from multiunit activity to select one out of these brain stimulation (DBS), a widely prescribed form of
six motions. The encoding scheme involved activating the neurostimulation for reducing tremors and restoring
movement associated with the highest decoder output motor function in Parkinson’s patients. Current DBS
using an electrode stimulation pattern previously cali- systems are open-loop but Herron et al. have recently
brated to evoke that movement. Surface electrical stimula- demonstrated closed-loop DBS [44] by triggering DBS
tion was delivered as monophasic rectangular pulses at based on movement intention, which was decoded as
50 Hz pulse rate and 500 ms pulse width, with stimulation reduction in ECoG power in the low frequency (‘mu’)
intensity set to a piecewise linear function of decoder band over motor cortex.
output. These results were extended to multi-joint reach-
ing and grasping movements by Ajiboye et al. [42

]: a linear Enhancing memory and augmenting brain function
decoder similar to a Kalman filter was used to map neuronal Besides restoration of lost function, BBCIs can also be
firing rates and high frequency power at electrodes in the used for augmenting brain function. Berger et al. [45

,46]
hand area of the motor cortex to percent activation of have demonstrated that BBCIs implanted in the hippo-
stimulation patterns associated with elbow, wrist or hand campus of monkeys and rats can be used to enhance
movements. The researchers showed that a tetraplegic memory in delayed match-to-sample (DMS) and non-
subject could perform multi-joint arm movements for match-to-sample tasks. They first fit a multi-input/multi-
point-to-point target acquisitions with 80–100% accuracy output (MIMO) nonlinear filtering model to
and volitionally reach and drink a mug of coffee. simultaneously recorded spiking data from hippocampal
CA3 and CA1 during successful trials, with CA3 as input
One shortcoming of the above approaches is that contin- to the model and CA1 as output. The trained MIMO
ued electrical stimulation of muscles results in muscle model was later used to decode CA3 activity and predict
fatigue, rendering the technique impractical for day-long CA1 activity encoded as patterns of biphasic electrical
use. An alternate approach to reanimation is to use brain pulses. Deadwyler et al. [46] showed that in the four
signals to stimulate the spinal cord. Spinal stimulation monkeys tested, performance in the DMS task was
may simplify encoding and control because it activates enhanced in the difficult trials, which had more distractor
functional synergies, reflex circuits, and endogenous pat- objects or required information to be held in memory for
tern generators. Capogrosso et al. [43] demonstrated the longer durations. However, it is unclear how the approach
efficacy of brain-controlled spinal stimulation for hind could be used when the brain is not healthy such as in
limb reanimation for locomotion in paralyzed monkeys. Alzheimer’s patients [47] where simultaneous recordings
They used a decoder based on linear discriminant from areas such as CA3 and CA1 for training the model in
analysis to predict foot-strike and foot-off events during successful trials will not be available.
locomotion. The encoder used this prediction to activate
extensor and flexor ‘hotspots’ in the lumbar spinal cord Nicolelis suggested several brain augmentation schemes
via epidural electrical stimulation to correctly produce the based on BBCIs in his book [48], including direct brain-to-
extension and flexion of the impaired leg. brain communication. Pais-Vieira et al. subsequently showed
how rats can use brain-to-brain interfaces (BBIs) to solve
Restoring motor and cognitive function sensorimotor tasks [49
]: an ‘encoder’ rat identified a stimulus
One of the early pioneers exploring bidirectional BCIs for and pressed one of two levers while its M1 cortex activity was
the restoration of brain function was Delgado [6
] who transmitted to the M1 cortex of a ‘decoder’ rat. The

Current Opinion in Neurobiology 2019, 55:142–151 www.sciencedirect.com

 Toward neural co-processors for the brain Rao 145

stimulation pattern was based on a Z score computed from already being used for medical applications, the vast
the difference in the number of spikes between the current majority of BCIs are still in their ‘laboratory testing’
trial and a template trial. If the decoder rat made the same phase. Consider, for example, the most commonly cited
choice as the encoder rat, both rats were rewarded for the BCI application of communication using brain signals
successful transfer of information between their two brains. alone. The maximum bit rate achieved by a human using
Rao et al. utilized noninvasive technologies to demonstrate an invasive BCI is currently about 3.7 bits/s and 39.2 cor-
the first human brain-to-brain interface [50
,51,52]. The rect characters per minute [16

]. This is an order of
intention of a ‘Sender’ who could perceive but not act was magnitude lower than average human typing speeds of
decoded from motor or visual cortex using EEG; this infor- about 150–200 characters per minute. In noninvasive
mation was delivered via transcranial magnetic stimulation EEG-based BCIs, the highest bit rate has been achieved
(TMS) to the motor or visual cortex of a ‘Receiver’ who could using steady state visually evoked potentials (SSVEP):
act but not perceive. The researchers showed that tasks such each choice on a menu is associated with a flickering
as a video game [50
] or ‘20 questions’ [52] could be com- visual stimulus (e.g. an LED) flashing at a specific known
pleted successfully through direct brain-to-brain collabora- frequency. SSVEP BCIs have achieved bit rates as high as
tion (see Refs. [53,54] for other examples). More recently, 5.3 bits/s or 60 characters per minute [59], which is again
brain-to-brain interfaces have been used to create a network an order of magnitude less than manual typing speeds,
of brains or a ‘BrainNet’ allowing groups of humans [55
] or with the added drawback of visual fatigue. BBCIs are
rats [56
] to solve tasks collaboratively. even more in their infancy. For example, the noninvasive
brain-to-brain interfaces in humans cited above have bit
Inducing plasticity and rewiring the brain rates of less than 1 bit/s, partly due to safety consider-
Hebb’s principle for plasticity states that connections ations of transcranial magnetic stimulation.
from a group A of neurons to a group B are strengthened
if A consistently fires before B, thereby strengthening Towards a unifying framework: neural co-
the causal relationship from A to B. Jackson et al. [57

] processors based on deep learning
demonstrated that such plasticity can be artificially A major limitation of current BBCIs is that they treat
induced in the motor cortex of freely behaving primates decoding and encoding as separate processes, and they do
by triggering stimulation at a site B a few milliseconds not co-adapt and jointly optimize a cost function with the
after a spike was detected at site A. After two days of nervous system. We propose that these limitations may be
continuous spike-triggered stimulation, the output gen- addressed using a ‘neural co-processor’ as shown in
erated by site A shifted to resemble the output from B, Figure 1. A neural co-processor uses two artificial neural
consistent with a strengthening of any weak synaptic networks, a co-processor network (CPN) and an emulator
connections that may have existed from neurons in A to network (EN), combined with a new type of deep
neurons in B. Such an approach could be potentially learning that approximates backpropagation through both
quite useful for neurorehabilitation by rewiring the brain biological and artificial networks.
for the restoration of motor function after traumatic brain
injury, stroke or neuropsychiatric disorders such as Suppose the goal is to restore movement in a stroke or
depression and PTSD. Along these lines, Guggenmos spinal cord injury (SCI) patient, for example, to enable
et al. [58
] have shown that the approach can be used to the hand to reach a target object (see Figure 1). The CPN
improve reaching and grasping functions in a rat after is a multi-layered recurrent neural network that maps
traumatic brain injury to the rat’s primary motor cortex neural activity patterns from a large number of electrodes
(caudal forelimb area). Their approach involves creating in areas A1, A2, and so on (e.g. movement intention areas
an artificial connection between the rat’s premotor cor- spared by the stroke) to appropriate stimulation patterns
tex (rostral forelimb area or RFA) and somatosensory in areas B1, B2, and so on (e.g. intact movement execu-
cortex S1 and for each spike detected by an electrode in tion areas in the cortex or spinal cord). When the subject
RFA, delivering an electric pulse to S1 after 7.5 ms. All of forms the intention to move the hand to a target (e.g.
these prior approaches have relied on 1-to-1 spike-to- during a rehabilitation session), the CPN maps the
stimulation-pulse protocols, leaving open the question of resulting neural activity pattern to a stimulation pattern.
how the approach can be generalized to the induction of Unfortunately, to train the CPN, we do not have a set of
goal-directed multi-electrode plasticity, a question we ‘target stimulation patterns’ that produce the intended
address with the concept of ‘neural co-processors’ below. movements. However, for any stimulation pattern, we
can compute the error between the resulting hand move-
From proof-of-concept to real-world applications ment and the target. How can this behavioral error be
Most of the BBCIs reviewed above (see Table 1) involved translated and backpropagated through the CPN to gen-
proof-of-concept demonstrations. An important question erate better stimulation patterns?
is: how close are we to real-world applications of BCIs?
While a small number of BCIs such as deep brain stimu- We propose the use of an emulator network (EN) that
lators and Neuropace’s RNS epilepsy control system are emulates the biological transformation between

www.sciencedirect.com Current Opinion in Neurobiology 2019, 55:142–151

146 Machine learning, big data, and neuroscience

Table 1

Summary and comparison of some notable BBCIs built so far

BBCI Input and output Decoding and encoding Achievements Limitations

O’Doherty et al. [36 ] Spikes from monkey M1 Unscented Kalman filter Simultaneous brain- Artificial tactile feedback
cortex and intracortical and biphasic pulse trains controlled cursor and limited to reward/no
microstimulation in S1 of different frequencies artificial tactile feedback reward information
cortex in monkeys
Flesher et al. [29] Spikes from human M1 Velocity-based linear Simultaneous brain- Simple force matching
cortex and intracortical decoder and linear controlled prosthetic task with only two levels
microstimulation in S1 encoding of torque to hand and force feedback (gentle or firm)
cortex pulse train amplitude in humans
Moritz et al. [39

] Spikes from monkey M1 Volitional control of firing Direct brain control of Simple flexion and
cortex and functional rate of single neuron and paralyzed muscles to extension movements of
electrical stimulation of linear encoder restore wrist movement the wrist only, muscle
muscles in monkeys fatigue with prolonged
use
Bouton et al. [41

] Multiunit activity from Support vector machines Direct brain control of Decoder based on
hand area of human M1 for classifying one of six forearm muscles for classification of six fixed
cortex and functional wrist/hand motions and hand/wrist control in motions, muscle fatigue
electrical stimulation of previously calibrated paralyzed human with prolonged use
paralyzed forearm stimulation for each
muscles motion
Ajiboye et al. [42

] Spikes and high Linear decoder and Direct brain control of Percent activation of
frequency power in hand percent activation of arm muscles for multi- fixed stimulation
area of human M1 cortex stimulation patterns joint movements and patterns, muscle fatigue
and functional electrical associated with hand, point-to-point target with prolonged use
stimulation of paralyzed wrist or elbow acquisitions in paralyzed
arm muscles movements human
Capogrosso et al. [43] Multiunit activity from leg Linear discriminant Direct brain control of the Simple two-state
area of monkey M1 analysis to predict foot spinal cord for restoring decoder and encoder
cortex and epidural strike/foot off and locomotion in paralyzed models, viability for
electrical stimulation of activation of spinal monkeys restoration of bipedal
the lumbar spinal cord hotspots for extension/ walking in humans yet to
flexion be demonstrated
Delgado [6
] Local field potentials in Decoder algorithm for First BBCI to control Results not consistent
monkey amygdala and detecting fast ‘spindle’ behavior and induce from subject to subject,
electrical stimulation in waves and an electrical neuroplasticity in animals did not generalize to
the reticular formation stimulation for each treating depression in
detection humans
Deadwyler et al. [46] Spikes from area CA3 in Multi-input/multi-output First demonstration of Applicability to memory
monkey hippocampus (MIMO) nonlinear filtering memory enhancement in restoration in Alzheimer’s
and electrical model to decode CA3 a short-term memory or other patients unclear
microstimulation in area activity and encode task in a monkey due to MIMO model
CA1 in hippocampus predicted CA1 activity as training algorithm
biphasic electrical pulses
Jackson et al. [57

] Spikes from a region of Single spike detection First demonstration of Single input/single output
monkey M1 cortex and and biphasic electrical Hebbian plasticity protocol, not designed
intracortical pulse for each spike induction using a BBCI in for multi-input/multi-
microstimulation of a detection a freely behaving monkey output goal-directed
different region of M1 plasticity induction
Guggenmos et al. [58
] Spikes from rat premotor Single spike detection in First demonstration of Single input/single output
cortex and intracortical premotor cortex and improved motor function protocol, plasticity
microstimulation of S1 electrical pulse after traumatic brain induction not geared
somatosensory cortex stimulation in S1 after injury in a rat using a toward optimizing
7.5 ms BBCI for plasticity behavioral or
induction rehabilitation metrics

stimulation patterns and behavioral output. The EN is a acts as a surrogate for the biological networks mediating
deep recurrent neural network whose weights can be the transformation between inputs in B1, B2, and so on
learned using standard backpropagation from a dataset and output behavior.
consisting of a large variety of stimulation (or neural
activity) patterns in areas B1, B2, and so on and the With the help of a trained EN, we can train the weights of the
resulting movements or behavior. After training, the EN CPN to produce the optimal stimulation patterns for

Current Opinion in Neurobiology 2019, 55:142–151 www.sciencedirect.com

 Toward neural co-processors for the brain Rao 147

Figure 1

Co-Processor Network
Backpropagation of Error

External Sensor or
Information Source External Actuator

Neural
Neural Emulator Network
Stimulation
Recordings
Backpropagation of Error

Behavioral
or Task Error

Preserved motor
information flow
Stimulated
action

Error

Current Opinion in Neurobiology

Neural co-processor for the brain for restoring and augmenting function.
A deep recurrent artificial network is used to map input neural activity patterns in one set of regions to output stimulation patterns in other regions
(‘co-processor network’ or CPN). The CPN’s weights are optimized to minimize brain-activity-based error (between stimulation patterns and target
neural activity patterns when known), or more generally, to minimize behavioral error or task error using another network, an emulator network.
The emulator network is also a deep recurrent network that is pre-trained through backpropagation to learn the biological transformation from
stimulation or neural activity patterns at the stimulation site to the resulting output behaviors. During CPN training, errors are backpropagated
through the emulator network to the CPN to adapt the CPN’s weights but not the emulator network’s weights. The trained CPN, thus, produces
optimal stimulation patterns that minimize behavioral error, thereby creating a goal-directed artificial information processing pathway between the
input and output regions. The CPN also promotes neuroplasticity between weakly connected regions, leading to neural augmentation or targeted
rehabilitation. External information from artificial sensors or other information sources can be integrated into the CPN’s information processing as
additional inputs to the neural network and outputs can be computed for external actuators as well. The example here shows the CPN creating a
new information processing pathway between prefrontal cortex and motor cortex, bypassing an intermediate area affected by brain injury or
stroke. The CPN is trained to transform movement intentions in the prefrontal cortex to appropriate movement-related stimulation patterns in the
motor cortex for the restoration of movement and rehabilitation.

minimizing behavioral error (e.g. error between current hand promotes neuroplasticity between connected brain
position and a target location). For each neural input pattern regions via Hebbian plasticity. Note that unlike previous
X (e.g. movement intention) that the subject produces in plasticity induction methods [57

,58
], the plasticity
areas A1, A2 and so on, the CPN produces an output induced spans multiple electrodes and is goal-directed
stimulation pattern Y in areas B1, B2 and so on, which results since the CPN is trained to minimize behavioral errors.
in a behavior or movement Z. After a sufficient amount of coupling between regions X
and Y, neurons in region X can be expected to automati-
The error E between actual movement Z and the cally recruit neurons in region Y to achieve a desired
intended movement target Z’ is first backpropagated response (such as a particular hand movement). As a
through the EN but without modifying its weights. We con- result, in some cases, the neural co-processor system
tinue to backpropagate the error through the CPN, this may eventually be no longer required after a period of
time modifying the CPN’s weights. In other words, the use and may be removed once function is restored or
behavioral error is backpropagated through a augmented to a satisfactory level.
concatenated CPN–EN network but only the CPN’s
weights are changed. This allows the CPN to progres- To illustrate the generality of the neural co-processor frame-
sively generate better stimulation patterns that enable the work beyond restoring motor function, consider a co-processor
brain to better achieve the target behavior, thereby for emotional well-being (e.g. to combat trauma, depression,
resulting in a co-adaptive BBCI. or stress). The emulator network could first be trained by
stimulating one or more emotion-regulating areas of the brain
Furthermore, by repeatedly pairing patterns of neural and noting the effect of stimulation on the subject’s emotional
inputs with patterns of output stimulation, the CPN state, as captured, for example, by a mood score based on a

www.sciencedirect.com Current Opinion in Neurobiology 2019, 55:142–151

148 Machine learning, big data, and neuroscience

questionnaire answered by the subject [60]. The CPN could hand and the target or a sensor worn on the hand may
then be trained to map emotional intentions or other brain indicate error in the force applied. Similarly, in speech
states to appropriate stimulation patterns that lead to a desired rehabilitation after stroke, a speech analysis system could
emotional state (e.g. less traumatic, stressful or depressed quantify the error between the generated speech and the
state). target speech pattern.

Another example is using a neural co-processor to create A bigger challenge is to train an EN to be a sufficiently
a sensory prosthesis that converts sensory stimuli from an accurate model of the transformation from stimulation
external sensor (Figure 1), such as a camera, microphone, patterns to behavioral output. It may be difficult to obtain
or even infrared or ultrasonic sensor, into stimulation a sufficient amount of data containing enough examples
patterns that take into account the ongoing dynamics of of how stimulation affects behavior. One possible solution
the brain. In this case, the CPN in Figure 1 takes as input is to record neural activities in regions that are causally
both external sensor information and current neural related or correlated with observed behavior and use
activity to generate an appropriate stimulation pattern these data to train the EN, under the assumption that
in the context of the current state of the brain. The stimulation patterns will approximate the neural activity
emulator network could be trained based on the sub- patterns. Another possible approach is to build the EN in
ject’s reports of perceptual states generated by a variety a modular fashion, starting from biological structures
of stimulation patterns. closest to the target behavior and going up the hierarchy,
for example, learning to emulate the transformation from
More generally, the external input to the CPN could be limb muscles to limb movements, spinal activity to mus-
from any information source, even the internet, allowing cle activity, and so on. Finally, one could combine the
the brain to request information via the external actuator above ideas with the concept of transfer learning using
component in Figure 1. The resulting information is networks trained across similar neural regions or even
conveyed via the CPN’s input channels and processed across subjects, and incorporate prior knowledge from
in the context of current brain activity. The emulator computational neuroscience models of the biological
network in this case would be trained in a manner similar system being emulated. Regardless of the training
to the sensory prosthesis example above to allow the CPN method used, we expect that the EN (and CPN) will
to convert abstract information (such as text) into appro- need to be regularly updated with new neural data as the
priate stimulation patterns that the user can understand. brain adapts to having the CPN as part of its information
processing loops.
Finally, we note that neural co-processors can be useful
tools for testing new computational models of brain and Conclusions
nervous system function [31]. Rather than using tradi- Traditionally, much of BCI research has focused on the
tional artificial neural networks in the CPN in Figure 1, problem of decoding, specifically, how can movement
one could use more realistic cortical models such as intention be extracted from noisy brain signals to control
networks of integrate-and-fire or Hodgkin–Huxley neu- prosthetic devices? More recently, there has been grow-
rons, along with biological learning rules such as spike- ing interest in ‘closing the loop’ using bidirectional BCIs
timing-dependent plasticity rather than backpropagation. (BBCIs) which incorporate sensory feedback, for exam-
A critical test for putative computational models of the ple, from artificial tactile sensors, via stimulation. The
nervous system would then be: can the model success- ability to simultaneously decode neural activity from one
fully interact with its neurobiological counterpart and be region and encode information to deliver via stimulation
eventually integrated within the nervous system’s infor- to another region confers on BBCIs tremendous versatil-
mation processing loops? ity. In this article, we have reviewed how BBCIs can be
used to not only control prosthetic arms with sensory
Challenges feedback but also (1) control paralyzed limbs using motor
A first challenge in realizing the above vision for neural intention signals from the brain, (2) restore and augment
co-processors is obtaining an error signal for training the cognitive function and memory, and (3) induce neuro-
two networks. In the simplest case, the error may simply plasticity for rehabilitation. Promising proof-of-concept
be a neural error signal: the goal is to drive neural activity results have been obtained in animal models and in some
in areas B1, B2, and so on toward known target neural cases, humans, but mostly under laboratory conditions.
activity patterns, and we can therefore train the CPN
directly to approximate these activity patterns without To transition to real-world conditions, BBCIs must co-
using an EN. However, we expect such scenarios to be adapt with the nervous system and jointly optimize
rare. In the more realistic case of restoring motor behavior, behavioral cost functions. We introduced the concept
such as in stroke rehabilitation where the goal is, for of a neural co-processor which uses artificial neural net-
example, to reach towards a target, a computer vision works to jointly optimize behavioral error functions with
system could be used to quantify the error between the biological neural networks. A trained emulator network is

Current Opinion in Neurobiology 2019, 55:142–151 www.sciencedirect.com

 Toward neural co-processors for the brain Rao 149

used as a surrogate for the biological network producing This book makes the case for bidirectional BCIs as therapeutic devices
and describes the first results obtained from a bidirectional BCI called the
the behavioral output. Behavioral errors are backpropa- stimoceiver that modulated neural activity in a monkey that led to a
gated through the emulator network to the co-processor change in its behavior.
network which adapts its weights to minimize errors and 7. Vidal JJ: Toward direct brain–computer communication. Annu
delivers optimal stimulation patterns for specific neural Rev Biophys Bioeng 1973, 2:157-180.

input patterns. We illustrated how a neural co-processor 8. Chapin JK, Moxon KA, Markowitz RS, Nicolelis MA: Real-time
control of a robot arm using simultaneously recorded neurons
could be used to improve motor function in a stroke or in the motor cortex. Nat Neurosci 1999, 2:664-670.
spinal cord injury patient. Such co-processors have not yet
9. Velliste M, Perel S, Spalding MC, Whitford AS, Schwartz AB:
been validated in animal models or humans, but if suc- Cortical control of a prosthetic arm for self-feeding. Nature
cessful, they could potentially be applied to modalities 2008, 453:1098-1101.
other than movement such as: 10. Hochberg LR, Bacher D, Jarosiewicz B, Masse NY, Simeral JD,

Vogel J, Haddadin S, Liu J, Cash SS, van der Smagt P, Donoghue JP:
Reach and grasp by people with tetraplegia using a neurally

Mapping inputs from one memory-related area to controlled robotic arm. Nature 2012, 485:372-375.
another to facilitate or restore access to particular The authors demonstrate that two people with tetraplegia can use an
invasive brain–computer interface based on motor cortex activity to
memories (e.g. in memory loss) or to unlearn traumatic control a robotic arm to perform three-dimensional reach and grasp
memories (e.g. in PTSD), movements, including drinking coffee from a bottle.

Mapping inputs from novel external sensors or one 11. Wolpaw JR, McFarland DJ, Neat GW, Forneris CA: An EEG-based
sensory area to another to restore or augment sensation brain–computer interface for cursor control.
Electroencephalogr Clin Neurophysiol 1991, 78:252-259.
and perception,

Connecting areas involved in emotion processing to 12. Serruya MD, Hatsopoulos NG, Paninski L, Fellows MR,
Donoghue JP: Instant neural control of a movement signal.
augment or rehabilitate emotional function, and Nature 2002, 416:141-142.

Augmenting the brain’s knowledge, skills, information 13. Wolpaw JR, McFarland DJ: Control of a two-dimensional
processing, and learning capabilities with deep artificial
movement signal by a noninvasive brain–computer interface
neural networks. in humans. Proc Natl Acad Sci U S A 2004, 101:17849-17854.
The authors demonstrate how a noninvasive BCI based on scalp-
recorded electroencephalographic (EEG) signals can achieve perfor-
mance in a 2D cursor control task that are competitive with certain
Conflict of interest statement invasive BCIs using an adaptive linear decoding scheme based on
mu-rhythm and beta-rhythm frequency bands.
Nothing declared.
14. Li Z, O’Doherty JE, Hanson TL, Lebedev MA, Henriquez CS,
Nicolelis MA: Unscented Kalman filter for brain–machine
Acknowledgements interfaces. PLoS One 2009, 4:e6243.
This work was supported by the National Science Foundation (EEC-
1028725 and 1630178), the National Institute of Mental Health (CRCNS/ 15. Gilja V, Pandarinath C, Blabe CH, Nuyujukian P, Simeral JD,
NIMH 1R01MH112166-01), and a grant from the W.M. Keck Sarma AA, Sorice BL, Perge JA, Jarosiewicz B, Hochberg LR,
Foundation. The author would like to thank Eb Fetz, Chet Moritz, Andrea Shenoy KV, Henderson JM: Clinical translation of a high-
Stocco, Jeff Ojemann, Steve Perlmutter, Dimi Gklezakos, Jon Mishler, performance neural prosthesis. Nat Med 2015, 21:1142-1145.
Richy Yun, David Caldwell, Jeneva Cronin, Nile Wilson and James Wu for
16. Pandarinath C, Nuyujukian P, Blabe CH, Sorice BL, Saab J,
discussions related to topics covered in this article.

Willett FR, Hochberg LR, Shenoy KV, Henderson JM: High
performance communication by people with paralysis using
References and recommended reading an intracortical brain–computer interface. eLife 2017, 6:e18554.
Combining a specially designed Kalman filter known as the recalibrated feed-
Papers of particular interest, published within the period of review,
have been highlighted as: back intention-trained Kalman filter (ReFIT-KF) for cursor control and a hidden
Markov model for clicking, the authors demonstrate the highest information rate

of special interest to date for communication via an invasive brain–computer interface.


of outstanding interest 17. Farwell LA, Donchin E: Talking off the top of your head: toward a
mental prosthesis utilizing event-related brain potentials.
1. Rao RPN: Brain–Computer Interfacing: An Introduction. New York, Electroencephalogr Clin Neurophysiol 1988, 70:510-523.
NY: Cambridge University Press; 2013.
18. Sellers EW, Kübler A, Donchin E: Brain–computer interface
2. Wolpaw J, Wolpaw EW (Eds): Brain–Computer Interfaces: research at the University of South Florida Cognitive
Principles and Practice. Oxford University Press; 2012. Psychophysiology Laboratory: the P300 speller. IEEE Trans
Neural Syst Rehabil Eng 2006, 14:221-224.
3. Moritz CT, Ruther P, Goering S, Stett A, Ball T, Burgard W,
Chudler EH, Rao RP: New perspectives on neuroengineering 19. Bell CJ, Shenoy P, Chalodhorn R, Rao RPN: Control of a
and neurotechnologies: NSF-DFG workshop report. IEEE Trans humanoid robot by a noninvasive brain–computer interface in
Biomed Eng 2016, 63:1354-1367. humans. J Neural Eng 2008, 5:214-220.
4. Lebedev MA, Nicolelis MA: Brain-machine interfaces: from 20. Galán F, Nuttin M, Lew E, Ferrez PW, Vanacker G, Philips J, del
basic science to neuroprostheses and neurorehabilitation. R. Millán J: A brain-actuated wheelchair: asynchronous and
Physiol Rev 2017, 97:767-837. non-invasive brain–computer interfaces for continuous
control of robots. Clin Neurophysiol 2008, 119:2159-2169.
5. Fetz EE: Operant conditioning of cortical unit activity. Science

1969, 163:955-958. 21. del R. Millán J, Galán F, Vanhooydonck D, Lew E, Philips J,
This article was the first to demonstrate a brain–computer interface in a Nuttin M: Asynchronous non-invasive brain-actuated control
monkey. The author presents results showing that neurons in the motor of an intelligent wheelchair. Conference of the IEEE Engineering
cortex can be trained to control an artificial device through operant in Medicine and Biology Society. 2009:3361-3364.
conditioning.
22. Bryan M, Nicoll G, Thomas V, Chung M, Smith JR, Rao RPN:
6. Delgado J: Physical Control of the Mind: Toward a Psychocivilized Automatic extraction of command hierarchies for adaptive

Society. New York: Harper and Row; 1969. brain–robot interfacing. Proceedings of ICRA 2012 2012.

www.sciencedirect.com Current Opinion in Neurobiology 2019, 55:142–151

150 Machine learning, big data, and neuroscience

23. Niparko J (Ed): Cochlear Implants: Principles and Practices, edn 2. 41. Bouton CE, Shaikhouni A, Annetta NV, Bockbrader MA,
Philadelphia: Lippincott; 2009.

Friedenberg DA, Nielson DM, Sharma G, Sederberg PB, Glenn BC,
Mysiw WJ et al.: Restoring cortical control of functional movement
24. Weiland JD, Liu W, Humayun MS: Retinal prosthesis. Annu Rev in a human with quadriplegia. Nature 2016, 533:247-250.
Biomed Eng 2005, 7:361-401. This work extends prior work done in monkeys on cortical control of
paralyzed muscles to a human with quadriplegia. Multiunit activity from
25. Tomlinson T, Miller LE: Toward a proprioceptive neural hand area of M1 cortex was classified into six wrist/hand motions using
interface that mimics natural cortical activity. Adv Exp Med Biol functional electrical stimulation of paralyzed forearm muscles.
2016, 957:367-388.
42. Ajiboye AB, Willett FR, Young DR, Memberg WD, Murphy BA,
26. Tabot GA, Dammann JF, Berg JA, Tenore FV, Boback JL,


Miller JP, Walter BL, Sweet JA, Hoyen HA, Keith MW et al.:
Vogelstein RJ, Bensmaia SJ: Restoring the sense of touch with
Restoration of reaching and grasping movements through
a prosthetic hand through a brain interface. Proc Natl Acad Sci
brain-controlled muscle stimulation in a person with
U S A 2013, 110:18279-18284.
tetraplegia: a proof-of-concept demonstration. Lancet 2017,
27. Tyler DJ: Neural interfaces for somatosensory feedback: 389:1821-1830.
bringing life to a prosthesis. Curr Opin Neurol 2015, 28:574-581. Neural activity from the hand area of human M1 cortex was mapped to
percent activation of functional electrical stimulation patterns applied to
28. Dadarlat MC, O’Doherty JE, Sabes PN: A learning-based paralyzed arm muscles to elicit hand, wrist or elbow movements and

approach to artificial sensory feedback leads to optimal achieve point-to-point target acquisitions in a tetraplegic human.
integration. Nat Neurosci 2015, 18:138-144.
The authors present a new learning-based approach to sensory encoding 43. Capogrosso M, Milekovic T, Borton D, Wagner F, Moraud EM,
for artificial sensory feedback in which two monkeys learned to extract Mignardot JB, Buse N, Gandar J, Barraud Q, Xing D et al.: A brain-
hand position information relative to an unseen target from intracortical spine interface alleviating gait deficits after spinal cord injury
microstimulation patterns delivered via 7–8 electrodes implanted in in primates. Nature 2016, 539:284-288.
primary somatosensory cortex.
44. Herron JA, Thompson MC, Brown T, Chizeck HJ, Ojemann JG,
29. Flesher SN, Collinger JL, Foldes ST, Weiss JM, Downey JE, Tyler- Ko AL: Cortical brain–computer interface for closed-loop deep
Kabara EC, Bensmaia SJ, Schwartz AB, Boninger ML, Gaunt RA: brain stimulation. IEEE Trans Neural Syst and Rehabil Eng 2017,
Intracortical microstimulation of human somatosensory 25:2180-2187.
cortex. Sci Transl Med 2016, 8:361ra141.
45. Berger T, Hampson R, Song D, Goonawardena A, Marmarelis V,
30. Cronin JA, Wu J, Collins KL, Sarma D, Rao RP, Ojemann JG,

Deadwyler S: A cortical neural prosthesis for restoring and
Olson JD: Task-specific somatosensory feedback via cortical enhancing memory. J Neural Eng 2011, 8:046017.
stimulation in humans. IEEE Trans Haptics 2016, 9:515-522. This paper shows that a bidirectional BCI can be used to map multi-
neuronal activity in area CA3 of the hippocampus to stimulation patterns
31. Wander JD, Rao RPN: Brain–computer interfaces: a powerful in area CA1 using a nonlinear model to enhance the performance of a rat
tool for scientific inquiry. Curr Opin Neurobiol 2014, 25:70-75. in a memory task.
32. Mena GE, Grosberg LE, Madugula S, Hottowy P, Litke A, 46. Deadwyler SA, Hampson RE, Song D, Opris I, Gerhardt GA,
Cunningham J, Chichilnisky EJ, Paninski L: Electrical stimulus Marmarelis VZ, Berger TW: A cognitive prosthesis for memory
artifact cancellation and neural spike detection on large multi- facilitation by closed-loop functional ensemble stimulation of
electrode arrays. PLoS Comput Biol 2017, 13:e1005842. hippocampal neurons in primate brain. Exp Neurol 2017,
287:452-460.
33. O’Shea DJ, Shenoy KV: ERAASR: an algorithm for removing
electrical stimulation artifacts from multielectrode array 47. Senova S, Chaillet A, Lozano AM: Fornical closed-loop stimulation
recordings. J Neural Eng 2018, 15:026020. for Alzheimer’s disease. Trends Neurosci 2018, 41:418-428.
34. Zhou A, Johnson BC, Muller R: Toward true closed-loop 48. Nicolelis MAL: Beyond Boundaries. Macmillan; 2011.
neuromodulation: artifact-free recording during stimulation.
Curr Opin Neurobiol 2018, 50:119-127. 49. Pais-Vieira M, Lebedev M, Kunicki C, Wang J, Nicolelis MA: A

brain-to-brain interface for real-time sharing of sensorimotor
35. O’Doherty JE, Lebedev MA, Hanson TL, Fitzsimmons NA, information. Sci Rep 2013, 3:1319.
Nicolelis MA: A brain–machine interface instructed by direct The article describes the first brain-to-brain interface in animals in which
intracortical microstimulation. Front Integr Neurosci 2009, 3:20. neural activity in M1 cortex of an ‘encoder’ rat was converted into
36. O’Doherty JE, Lebedev MA, Ifft PJ, Zhuang KZ, Shokur S, intracortical microstimulation in M1 cortex of a ‘decoder’ rat, allowing


Bleuler H, Nicolelis MA: Active tactile exploration using a brain– the rats to achieve a common behavioral goal.
machine–brain interface. Nature 2011, 479:228-231.
50. Rao RP, Stocco A, Bryan M, Sarma D, Youngquist TM, Wu J,
This article presents results from two monkeys who were able to perform

Prat CS: A direct brain-to-brain interface in humans. PLoS One
tactile exploration of virtual objects with different artificial tactile proper-
2014, 9:e111332.
ties using a virtual arm controlled by motor cortex activity, while receiving
The article presents the first noninvasive brain-to-brain interface in
feedback via intracortical microstimulation in the somatosensory cortex.
humans. Motor imagery was decoded from EEG signals recorded from
37. Klaes C, Shi Y, Kellis S, Minxha J, Revechkis B, Andersen RA: A the motor cortex of one subject and used to deliver information to the
cognitive neuroprosthetic that uses cortical stimulation for hand area of the motor cortex of another subject through transcranial
somatosensory feedback. J Neural Eng 2014, 11:056024. magnetic stimulation (TMS), allowing the two subjects to collaborate and
solve a visual-motor task.
38. Flesher S et al.: Intracortical microstimulation as a feedback


source for brain–computer interface users. Brain–Computer 51. Rao RPN, Stocco A: When two brains connect. Sci Am Mind
Interface Research. 2017:43-54. 2014, 25:36-39.
The article presents first results from patients who are able to use artificial
sensory feedback provided through intracortical microstimulation in S1 52. Stocco A, Prat CS, Losey DM, Cronin JA, Wu J, Abernethy JA,
cortex to solve a force-feedback task involving a prosthetic hand con- Rao RPN: Playing 20 questions with the mind: collaborative
trolled by M1 cortex activity. problem solving by humans using a brain-to-brain interface.
PLoS One 2015, 10:e0137303.
39. Moritz CT, Perlmutter SI, Fetz EE: Direct control of paralysed


muscles by cortical neurons. Nature 2008, 456:639-642. 53. Grau C, Ginhoux R, Riera A, Nguyen TL, Chauvat H, Berg M,
The authors show that monkeys can be operantly conditioned to mod- Amengual JL, Pascual-Leone A, Ruffini G: Conscious brain-to-
ulate the firing rates of motor cortical neurons to control the electrical brain communication in humans using non-invasive
stimulation of paralyzed muscles to move the wrist and solve a motor technologies. PLoS One 2014, 9:e105225.
task.
54. Lee W, Kim S, Kim B, Lee C, Chung YA, Kim L, Yoo SS: Non-
40. Ethier C, Oby ER, Bauman MJ, Miller LE: Restoration of grasp invasive transmission of sensorimotor information in humans
following paralysis through brain-controlled stimulation of using an EEG/focused ultrasound brain-to-brain interface.
muscles. Nature 2012, 485:368-371. PLoS One 2017, 12:e0178476.

Current Opinion in Neurobiology 2019, 55:142–151 www.sciencedirect.com

 Toward neural co-processors for the brain Rao 151

55. Jiang L, Stocco A, Losey DM, Abernethy JA, Prat CS, Rao RPN: The authors present the first demonstration of cortical plasticity

BrainNet: a multi-person brain-to-brain interface for direct induction in M1 cortex based on activity-dependent electrical stimu-
collaboration between brains. Nat Sci Rep 2019. lation in freely behaving monkeys using a bidirectional brain–computer
arXiv:1809.08632, in press. interface.
The article presents results from a noninvasive brain-to-brain interface that
decodes EEG from the visual cortices of two subjects and transmits the 58. Guggenmos DJ, Azin M, Barbay S, Mahnken JD, Dunham C,
information via TMS to the visual cortex of a third subject. The authors show
Mohseni P, Nudo RJ: Restoration of function after brain
that this ‘BrainNet’ interface allows more than two brains to collaboratively damage using a neural prosthesis. Proc Natl Acad Sci U S A
solve a task that none of the individual brains could solve on their own. 2013, 110:21177-21182.
The authors show that activity-dependent stimulation using a bidirec-
56. Pais-Vieira M, Chiuffa G, Lebedev M, Yadav A, Nicolelis MA: tional BCI can be used to induce cortical plasticity in rats to improve

Building an organic computing device with multiple reaching and grasping functions after traumatic brain injury to the rat’s
interconnected brains. Sci Rep 2015, 5:11869. primary motor cortex (caudal forelimb area).
The article presents the first demonstration of a ‘Brainet’ of four inter-
connected rat brains with electrode arrays implanted bilaterally in S1 59. Chen X, Wang Y, Nakanishi M, Gao X, Jung TP, Gao S: High-
cortex for recording and stimulation. The Brainet functioned as an organic speed spelling with a noninvasive brain–computer interface.
computing device that could be used to solve problems involving discrete Proc Natl Acad Sci U S A 2015, 112:E6058-E6067.
classification, sequential and parallel computations, and memory sto-
rage/retrieval. 60. Sani OG, Yang Y, Lee MB, Dawes HE, Chang EF,
Shanechi MM: Mood variations decoded from multi-site
57. Jackson A, Mavoori J, Fetz EE: Long-term motor cortex plasticity intracranial human brain activity. Nat Biotechnol 2018,


induced by an electronic neural implant. Nature 2006, 444:56-60. 36:954-961.

www.sciencedirect.com Current Opinion in Neurobiology 2019, 55:142–151