KR20190051766A - Neuron Circuit, system and method for synapse weight learning - Google Patents

Abstract

Provided is a neuron circuit executing synapse weight learning for a plurality of synapse weight values. More specifically, the neuron circuit may determine whether a received input signal is an activation signal, whose synapse value is an active value, perform reinforcement learning setting a synapse weight value of a previously received input signal according to first probability as the active value, and perform weakening learning to set a plurality of synapse weight values as inactive values according to second probability.

Description

{Neuron Circuit, system and method for synapse weight learning}

The present disclosure relates to neuron circuits, systems and methods for synaptic weighted learning.

An artificial neural network refers to an artificial neuron or a computing device for implementing the interconnections of the models.

Convolution Neural Network and Recursive Neural Network have been proposed as an implementation method of artificial neural network. Furthermore, spiking artificial neural network, which imitates human biological structure and constructs artificial neural network similar to human brain structure, Research on Spiking Neural Network (SNN) has been highlighted.

Unlike the conventional method, SNN is optimized for learning dynamic features, and software and hardware methods for implementing it are proposed.

However, in the conventional method, the learning speed is limited due to a large amount of computation processing, and the learning speed is gradually slowed down to the upper layer. Also, according to the existing deterministic method, a large amount of memory There is a problem in that the system cost increases.

This disclosure relates to neural circuits, systems and methods for synaptic weighted learning in accordance with various embodiments, and proposes a method for facilitating discrete implementation using a probabilistic learning method and improving learning speed.

A neuron circuit for performing a synapse learning on a plurality of synapse weight values according to an aspect comprises a synapse weight memory for storing a plurality of synapse weight values, a receiving circuit for receiving an input signal from a pre- A second subcircuit for performing a comparison operation between a first cumulative number of times of reception of the activation input signal and a learning threshold value in accordance with a determination result, And performing reinforcement learning to set the synaptic weight value of the at least one input signal previously received according to the first probability to an active value when the first cumulative number of received times reaches a learning threshold, In accordance with a second probability, to an inactive value.

The learning system for a plurality of synapse weight values according to another aspect may include a neuron circuit and a learning circuit.

Further, the neuron circuit receives the input signal from the synapse pre-neuron circuit, and judges whether or not the received input signal is the activation signal whose synapse value of the received input signal is the active value, so that the first accumulated reception number of the activated input signal is When the learning threshold value is reached, the learning request signal can be transmitted to the learning circuit.

The learning circuit also performs reinforcement learning with a first probability of setting the synaptic weight value of the at least one input signal previously received by the neuron circuit to the active value as it receives the learning request signal, It is possible to perform weak learning in which each of the weight values is set to an inactive value according to the second probability.

A learning method for a plurality of synapse weight values between a neuron circuit and a synapse neuron circuit, which is performed by a neuron circuit according to another aspect, includes receiving, as an input signal from a synapse pre-neuron circuit, Determining whether a synapse value of the received input signal is an active input signal that is an active value, performing a comparison operation between a first cumulative number of times of reception of the activation input signal and a learning threshold value according to a determination result, Performing a reinforcement learning step of setting a synapse weight value of at least one input signal that has been previously received to an active value according to a first probability when the number of times reaches a learning threshold, , And setting an inactive value according to the weakened learning step.

There is provided a computer-readable recording medium recording a program for causing a computer to execute a learning method for a plurality of synapse weight values between a neuron circuit and a synapse pre-neuron circuit according to another aspect.

According to the embodiments, since the neuron circuit can perform spiked reinforcement learning and weakening learning, a spiking artificial neural network can be implemented using a synapse weight having a binary value, thereby reducing the memory capacity for storing the synaptic weight have.

In addition, the neuron circuit can prevent the learning rate from being lowered in the upper layer due to the lower learning speed of the lower layer by separating the learning process and the speech process.

FIGS. 1A to 1C are schematic views for explaining a spiking neural network according to an embodiment.
FIGS. 2A through 2D are diagrams illustrating a time-dependent spike-timing-dependent plasticity (STDP) learning function according to an embodiment.
3A through 3D are diagrams illustrating an order dependent STDP learning function according to an embodiment.
4 is a diagram illustrating a kernel function according to an embodiment.
5 is a diagram illustrating a hierarchical population of SNNs according to one embodiment.
6 is a block diagram of a neuron circuit according to one embodiment.
7 is a block diagram of a neuron circuit according to one embodiment.
8 is a diagram illustrating a synapse weight learning system according to an embodiment.
9 is a diagram illustrating a synapse weight learning system for a plurality of neuron circuits according to one embodiment.
10 is a diagram showing a learning system of a hierarchical neuron circuit group.
11 is a flow diagram of a method of learning a synaptic weight performed by a neuron circuit according to one embodiment.
12 is a diagram showing a scheme of a neuron block for implementing a neuron circuit.
13 is a diagram illustrating an STDP learning block scheme for implementing a learning circuit according to an embodiment.
14 is a diagram illustrating a single core block scheme for implementing a single core circuit including a plurality of neuron circuits according to one embodiment.
15 is a diagram showing a multicore block including a plurality of single core blocks.

FIGS. 1A through 1C are schematic views for explaining a Spiking Neural Network (SNN) according to an embodiment of the present invention.

SNN is an artificial neural network that imitates a biological neural network mechanism that includes neurons in the brain. The SNN operates based on the generation of discrete spike signals.

를 가질 수 있다. 이 때, i는 뉴런 각각에 할당된 정수 인덱스를 나타낼 수 있다. Referring to FIG. 1A, the neuron 10 includes an input terminal into which an input spike signal is input. The received input spike signal may have different properties and shapes depending on the physical or computational implementation. If the SNN is a purely computationally implemented implementation, such as a software program, the input spike signal can be represented by the time information the corresponding input spike signal was received. Alternatively, the input spike signal may change over time and be represented by a complex nerve spike signal having a biological form. When a neuron is physically implemented, the input spike signal may be a pulse of current, voltage, charge, magnetism, or a more complex and sophisticated time-dependent signal. Generally, a neuron has at least one internal state variable

Lt; / RTI > In this case, i may represent an integer index assigned to each neuron.

를 증가 시키거나, 감소시킬 수 있다. 전자를 양의 기여도라 하며 후자를 음의 기여도라 한다. 이는 뉴런이 수신된 입력 스파이크 신호를 통합(integrating)한다고 표현된다. 또한 누출 메커니즘(leakage mechanism)에 따라, 스파이크 신호를 입력 받는 사이에 뉴런의 상태 변수

는 휴지 상태 값으로 수렴하는 경향을 갖는다.　누출 메커니즘에 따라, 뉴런이 스파이크 신호를 충분히 오랜 시간 수신하지 않으면, 뉴런은 휴지 상태에 도달한다.The input spike signal is a state variable

Can be increased or decreased. The former is called the positive contribution, and the latter is the negative contribution. This is expressed as a neuron integrating the received input spike signal. In addition, depending on the leakage mechanism, the state variable of the neuron during the input of the spike signal

Has a tendency to converge to a dormant state value. Depending on the leak mechanism, if a neuron does not receive the spike signal long enough, the neuron will reach a dormant state.

FIG. 1B shows the interconnection relationship between neurons according to an embodiment.

When implemented biologically, artificially, physically, and discrete, neurons are interconnected via "synapses." Synapses are unidirectional. That is, the signal is transmitted from the pre-synaptic terminal to the post-synaptic terminal. Synapse adjusts the magnitude of the effect of the output spike signal of the synaptic neuron that transmits the signal on the input spike signal of the post-synaptic neuron receiving the signal.

에 의해 결정된다. 시냅스의 가중치

는 시냅스 전 뉴런을 나타내는 인덱스 i와 시냅스 후 뉴런을 나타내는 인덱스 j의 조합으로 표현된다.The degree of coordination is generally determined by the weight of the synapse

. Weight of synapse

Is represented by a combination of an index i representing a pre-synapse neuron and an index j representing a post-synapse neuron.

The weights of synapses can change over time based on "learning rules". For example, an artificial neural network can perform learning based on a pattern of a spike signal input to an artificial neural network, a spike signal generated in an artificial neural network, or the like.

In the case of supervised learning, the external monitoring agent participates in whether to learn a given target function. In unsupervised learning, there is no external monitoring agent involved in learning. In non - articulation learning, artificial neural networks learn statistical representation of input features. STDP learning is an example of non-learning of SNN.

1C is a graph illustrating a STDP learning function according to an embodiment.

는 하기 수학식 1에 따른 시냅스 후 스파이크의 발생 시간

및 시냅스 전 스파이크의 발생 시간

의 차이인

에 의해 결정된다.For synapses that connect the synaptic neuron i and post-synaptic neuron j, the weight of the synapse

Is the time of occurrence of the post-synaptic spike according to the following formula

And time of occurrence of synaptic spikes

Difference of

.

가 양수로서, 시냅스 전 스파이크와 시냅스 후 스파이크가 인과관계인 경우, 해당 시냅스 가중치가 강화되고, 그렇지 않은 경우 해당 시냅스 가중치가 약화된다.In Equation (1)

Is positive, if the synaptic spike and post-synaptic spike are causal, the corresponding synapse weights are enhanced, otherwise the corresponding synapse weights are weakened.

In general, the value of the STDP learning function is not 0 in the limited time window [-Tmin, + Tmax]. For example, biologically Tmin and Tmax have values in the range of 100 ms. However, in the artificial SNN, Tmin and Tmax vary based on the dynamic time constant of the pattern to be learned.

Figures 2A-2D illustrate a time dependent STDP learning function according to one embodiment.

2A is a graph showing a graph of a causal STDP learning function according to an embodiment.

If the ratio of the A + area to the A-area is similar to the ratio between the A + area and the A-area corresponding to Fig. 1c, similar results are obtained in synaptic strengthening and weakening.

2B is a graph showing a graph of a narrow potentiation window STDP learning function.

Reinforcement learning is performed only in a narrow range of positive time windows, [0, Tp], and weakening learning is performed in other time zones of the entire time window [-Tmin, Tmax] .

At this time, when Tmax expands to positive infinity and -Tmin expands to negative infinity, all of the synapses that connect the spiked target neurons to the target neuron's synaptic neurons are weakened by a fixed amount, Net potentiation occurs at the synapse that transmits a spike signal from the pre-synaptic preneuron to the target neuron within a predetermined time before the spike of the spike occurs. This is defined as "indiscriminate depressing STDP".

Figure 2c is a graph of a STDP learning function that defines only symmetric hebbian potentiation.

Referring to FIG. 2C, when the difference between the generation time of the synapse pre-spike signal and the generation time of the post-synapse spike signal is within a predetermined time regardless of the pre-synapse, synapse strengthening can be performed.

2D is a graph showing a symmetric heavyian STDP learning function with a narrow window.

2d, reinforcement learning occurs when the difference between the generation time of the synapse pre-spike signal and the generation time of the post-synapse spike signal belongs to the limited time window [-Tn, Tp], and weakening occurs in the remaining area do.

At this time, when Tmax expands to positive infinity, and -Tmin expands to negative infinity, learning according to the STDP function of FIG. 2D is defined as " symmetric indiscriminate weakening STDP ".

3A through 3D are diagrams illustrating an order dependent STDP learning function according to an embodiment.

As described above, the figures shown in Figs. 2A to 2D are time dependent learning functions. Thus, in implementing a hardware or software algorithm, a timestamp, which is information indicating the spike occurrence time of the synaptic neuron or the spike occurrence time of the post-synapse neuron, is required.

Correspondingly, order dependent learning functions are proposed. Order dependent learning functions can be obtained by replacing the time axis of time dependent learning functions in a discrete order. Thus, STDP learning with order dependent learning functions can be implemented by tracking the order of pre-synaptic events where spikes of the pre-synaptic neurons occur and post-synaptic events where spikes of post-synaptic neurons occur. For example, an event list in which events of the entire artificial neural network system are arranged in order can be stored integrally, or the event list of each of the subsystems can be separately stored. Alternatively, an event list for each synapse may be separately stored. In storing the event list, by setting the maximum number N of storable events occurring in the latest order, the neuron network can be implemented to perform the learning using the events that occurred recently.

대신, 하기 수학식 2에 따른 시냅스 후 뉴런 스파이크의 순서 인덱스

및 시냅스 전 뉴런 스파이크의 순서 인덱스

간의 인덱스 차이인

으로 치환된 점에서 차이가 있다.Figures 3a-3d are sequence dependent learning functions corresponding to Figures 2a-2d, in which the horizontal axis represents the time of post-synaptic spikes and the time difference between signals of the synaptic spikes

Instead, the order index of post-synaptic neuron spikes according to Equation (2)

And order index of synaptic neuronal spikes

Index difference between

. In this case,

The feature of sequence - based STDP learning is that learning is adaptive to the dynamics of neural activity. Because STDP is a time-based learning method, this is fundamentally different from biological STDP, which depends on a range of time constants and dynamics.

Order-based STDP learning is performed independently of the rate at which events occur and dynamics. Even if the timestamps of all spikes are multiplied by the same constant, the order is the same, so the learning content is not changed as a result. Therefore, sequence based STDP has its own adaptability according to the speed and dynamics of events.

/w 또는

를 이전 가중치에 적용한다. 이 때, 스파이크 시퀀스는 특정 선호 패턴에 치우치는 것을 막기 위해 자극(stimuli) 순서를 변경하면서 반복된다.In general, in the computational implementation of STDP, the learning function is characterized by a small amount of deterministic synaptic variation, for example, 1% or 0.1%

/ w or

To the previous weight. At this time, the spike sequence is repeated while changing the stimuli order to prevent bias to a certain preference pattern.

Thus, a typical STDP is deterministic and requires a high resolution to store the weights so that small changes in synaptic weights can be realized.

On the other hand, applying the probabilistic nature to the learning process can lower the resolution needed for storing synaptic weights. This can be implemented by applying a high variation amount with a small probability according to the probabilistic learning method, instead of applying a small variation amount to the weight according to the deterministic learning method.

을 미리 결정된 양만큼 가중치 변화가 발생할 확률 p로 치환함으로써 구현될 수 있다.For example, a probabilistic STDP learning model can be implemented using the learning functions of FIGS. 2A through 2D and FIGS. 3A through 3D. This is because the weight change amount

With a probability p that a weight change will occur by a predetermined amount.

또는

이 양의 확률 p (0<= p <=1)인 경우, 랜덤 수 x (0<=x<=1)를 생성하여, x<=p인 경우 가중치를 1로 설정하고, x>p인 경우 가중치를 유지할 수 있다. 만약, ξ(

t) 또는 ξ(

n)이 음의 확률 p (0>= p >=-1)인 경우, 랜덤 수 x (0<=x<=1)를 생성하여, x<=|p|인 경우 가중치를 0으로 설정하고, x>|p|인 경우 가중치를 유지할 수 있다.For example, the synapse weight value may be 1 or 0, which is a value of 1 bit. if,

or

If x <= p, the weight is set to 1, and if x> p (= 1), a random number x (0 <= x <= 1) If we can keep the weight. If ξ (

t) or ξ (

(0 <= x <= 1), if the probability of the negative (p) is 0 (p> 0), then the weight is set to 0 , x> | p |, the weight can be maintained.

4 is a diagram illustrating a kernel function according to an embodiment.

The kernel function represents the effect of the input signal input through the synapse on the internal state variables of the neuron. For example, biologically, the internal state variable of a neuron can indicate the membrane potential of a biological neuron.

If the value of the internal state variable of the neuron exceeds the threshold value, the neuron can fire and output the spike signal.

는 시간에 따른 함수로서 시간 영역에서 하기 수학식 3에 따른 미분 방정식을 만족한다.For example, the state variables of neurons

Satisfies the differential equation according to Equation (3) in the time domain as a function of time.

는 입력단, 또는 시냅스 전 뉴런에 대응하는 인덱스

및 현재 뉴런, 즉 시냅스 후 뉴런에 대응하는 인덱스 k에 의해 정의되는 두 뉴런간의 시냅스 가중치를 나타낸다.

는 휴지 상태의 상태 변수 값으로서 생물학적으로 뉴런 막 전위의 휴지 전압을 나타낼 수 있다. 우변의

텀은 시간 상수

에 의해 시간에 따라 감소하는 누출 메커니즘(leakage mechanism)을 나타낸다. 누출 메커니즘에 의해 뉴런이 장기간 입력 신호를 수신하지 않으면 뉴런의 상태 변수 값은 휴지 상태 값

으로 수렴한다. 누출 메커니즘은 exponential한 감쇄이나 하드웨어적 구현에서 선형 감쇄로 구현될 수 있다.In differential equations

Is an input, or an index corresponding to the pre-synaptic neuron

And the synapse weights between the two neurons defined by the current neuron, i.e., the index k corresponding to the post-synaptic neuron.

May be biased to indicate a resting voltage of the neuronal membrane potential as a state variable of the dormant state. Right

Term is a time constant

Lt; RTI ID = 0.0 &gt; leakage &lt; / RTI &gt; mechanism with time. If the neuron does not receive the input signal for a long time by the leakage mechanism, the value of the state variable of the neuron becomes the dormant state value

. The leakage mechanism can be implemented with either exponential attenuation or linear attenuation in hardware implementations.

에 미치는 영향을 나타내는 함수의 일 예인 커넬 함수 h(t)는 도 4에서와 같이 입력 시간 후 소정의 시간이 경과하기 까지 상태 변수에 미치는 기여도가 증가하다가 이후 점차 기여도가 감소하는 형태일 수 있다.On the other hand, as the input signal changes with time,

The kernel function h (t), which is an example of a function indicating the influence on the state variable, may be a form in which the contribution to the state variable increases until the predetermined time elapses after the input time as shown in FIG. 4, and then the contribution decreases gradually.

만큼 증가된다.The function h (t) can be simplified to a delta function for the efficiency of the computational implementation. Therefore, at the time when the input signal is input, the state variable immediately changes to the synapse weight

.

5 is a diagram illustrating a one-tier population of SNNs according to one embodiment.

는 출력 채널 O를 통하여 상위 계층인 P+1 계층 집단에 입력 신호로서 전송된다.In Figure 5, the P-tier population (P is an integer) of SNNs comprises a plurality of neurons. For example, each of the K neurons may be represented by an index k (k = 1, 2, ..., K). The signal output from each of the plurality of neurons

Is transmitted as an input signal to the P + 1 hierarchy group which is the upper layer through the output channel O. [

는 입력 채널 I를 통하여 입력단 i에서 수신된다.Also, the P-tier cluster of SNNs includes a plurality of input stages. For example, m input stages can be represented by index i (i = 1, 2, ..., m). Input signal

Is received at input i via input channel I.

, 입력 단을

라 할 때, 입력 신호를

로 나타낼 수 있다.In addition, when an input event occurs in which an input signal is received by the hierarchical group P,

, The input stage

, The input signal

.

, 스파이크 신호를 출력한 뉴런의 인덱스를

라 할 때, 출력 신호를

로 나타낼 수 있다.Similarly, when an output event occurs in which at least one of the neurons of the hierarchical group P outputs a spike signal, the output time

, The index of the neuron outputting the spike signal

, The output signal is

.

For example, each SNN layer may include an event store.

Each time an input event occurs, the event storage unit acquires information about the input event and constructs a pre-synaptic event list according to the time expressed by Equation (4).

Further, the event storage unit may acquire information about the output event every time an output event occurs, and construct a post-synapse event list expressed by the following Equation (5).

가 발생할 때마다, SNN 계층은 시냅스 후 이벤트 리스트의 출력 이벤트

각각에 대하여,

를 산출하여,

를 획득한다.Input event

, The SNN layer is responsible for the output event of the post-synaptic event list

For each,

Respectively,

.

가 0보다 큰 경우

는 시냅스 가중치

를 1로 설정할 확률이고, 0보다 작은 경우

는 시냅스 가중치

를 0으로 설정할 확률을 나타낸다.if,

Is greater than zero

Synaptic weights

Is set to 1, and if it is smaller than 0

Synaptic weights

Is set to zero.

가 발생할 때마다 SNN 계층은 시냅스 후 이벤트 리스트의 입력 이벤트

각각에 대하여,

를 산출하여,

를 획득한다.Output events

The SNN hierarchy is used to determine the input event of the post-synaptic event list

For each,

Respectively,

.

가 0보다 큰 경우

는 시냅스 가중치

를 1로 설정할 확률이고, 0보다 작은 경우

는 시냅스 가중치

를 0으로 설정할 확률을 나타낸다.if,

Is greater than zero

Synaptic weights

Is set to 1, and if it is smaller than 0

Synaptic weights

Is set to zero.

중 강화 학습 영역, 즉

가

인 범위에 속하는지 여부만 판단함으로써, STDP 학습이 수행될 수 있다.In the case of the indiscriminate weakened STDP learning function described in connection with Figure 2b, a post-synaptic event list may not be needed. Every time an output event occurs, the x-axis of Fig.

A reinforcement learning area, namely

end

, The STDP learning can be performed.

In case of sequence-based STDP, STDP learning can be performed only in the order of the pre-synapse event list and the post-synapse event list, without any additional information on the time.

For example, the pre-synapse event list can be represented by an index according to an input stage and a sequence as shown in Equation (6).

For example, the post-synapse event list can be represented by a synapse outputting a spike signal and an index according to a sequence as shown in Equation (7).

가 발생할 때마다 SNN 계층은 시냅스 후 이벤트 리스트의 출력 이벤트

각각에 대하여,

를 산출하여,

를 획득한다.Input event

The SNN layer is used to generate an output event of the post-synapse event list

For each,

Respectively,

.

가 0보다 큰 경우

는 시냅스 가중치

를 1로 설정할 확률이고, 0보다 작은 경우

는 시냅스 가중치

를 0으로 설정할 확률을 나타낸다.if,

Is greater than zero

Synaptic weights

Is set to 1, and if it is smaller than 0

Synaptic weights

Is set to zero.

가 발생할 때마다 SNN 계층은 시냅스 후 이벤트 리스트의 입력 이벤트

각각에 대하여,

를 산출하여,

를 획득한다.Output events

The SNN hierarchy is used to determine the input event of the post-synaptic event list

For each,

Respectively,

.

가 0보다 큰 경우

는 시냅스 가중치

를 1로 설정할 확률이고, 0보다 작은 경우

는 시냅스 가중치

를 0으로 설정할 확률을 나타낸다.if,

Is greater than zero

Synaptic weights

Is set to 1, and if it is smaller than 0

Synaptic weights

Is set to zero.

중 강화 학습 영역, 즉

이

인 범위에 속하는지 여부만 판단함으로써, STDP 학습이 수행될 수 있다.In the case of the indiscriminate weakened STDP learning function described in connection with Figure 3b, a post-synaptic event list may not be needed. Every time an output event occurs, the x-axis in Fig. 3

A reinforcement learning area, namely

this

, The STDP learning can be performed.

6 is a block diagram of a neuron circuit 600 according to one embodiment.

6, the neuron circuit 600 may include a synapse weight memory 610, a first subcircuit 620, a second subcircuit 630, and a third subcircuit 640. It is to be understood by those skilled in the art that only the essential components associated with the embodiment are shown in FIG. 6, and that other general components may be included.

The synapse weight memory 610 may store a plurality of synapse weight values. For example, the synapse weight value may be one bit of data. The synapse weight value may have an active value or a binary value of an inactive value. For example, the active value may be 1 and the inactive value may be 0, but is not limited thereto.

A synapse whose synaptic weight value is the active value is defined as an activated synapse, and an inactive value is defined as a deactivated synapse.

The synapse weight memory 610 may store a synapse weight value corresponding to each of a plurality of synapse pre-neuron circuits. For example, the synaptic weight memory 610 may store a synaptic weight value of each of a plurality of synapse pre-neuron circuits in each of a plurality of cells.

The synapse weight memory 610 may be accessible from other components or external components within the neuron circuit 600. For example, as the cell address information and the read operation request signal are applied, the synapse weight memory 610 may output a synapse weight value corresponding to the cell address information. Accordingly, the component transmitting the read operation request signal can obtain the synapse weight value corresponding to the cell address information. In addition, as the cell address information, the write operation request signal, and the weight value are applied to the synapse weight memory 610, the synapse weight value corresponding to the cell address information in the synapse weight memory 610 may be set to an authorized weight value have.

The first sub-circuit 620 receives an input signal from the synapse pre-neuron circuit and can determine whether the received input signal is an activation signal whose synaptic weight value of the received input signal is the active value.

For example, the first subcircuit 620 may receive an input signal. The input signal may be a signal output from the synaptic neuron circuit of the neuron circuit 600. For example, the input signal may be a spike signal output from a synaptic neuron circuit.

In addition, the input signal may be a signal output from a plurality of synaptic pre-neuron circuits. For example, the input signal may be a signal to which spike signals output from at least some of the plurality of synaptic pre-neuron circuits are input selectively or in combination.

For example, the input signal may include information that can identify a synaptic neuronal circuit that has output a spike signal. As another example, the input signal may include information indicating that the synapse pre-neuron circuit has output a spike signal and / or identification information of the synapse pre-neuron circuit. For example, the identification information contained in the input signal may include an identity unique to the synaptic neuron circuit. In addition, the identification information included in the input signal may include cell address information in the synaptic weight memory 610 in which the synapse weight value of the synapse pre-neuron circuit is stored.

The first subcircuit 620 may determine whether the input signal is an active input signal whose synaptic weight value corresponding to the input signal is the active value. For example, the first sub-circuit 620 may obtain a synapse weight value corresponding to the input signal from the synapse weight memory 610. For example, the first subcircuit 620 may apply the read operation request signal and the synapse cell address information included in the input signal to the synapse weight memory 610. In addition, the first sub-circuit 620 may obtain a synapse weight value corresponding to the synapse cell address information from the synapse weight memory 610. If the obtained synapse weight value is the active value, the first sub circuit 620 may determine that the input signal is an activation signal. If the obtained synapse weight value is an inactive value, the first sub circuit 620 can determine that the input signal is not the activation signal. The first sub circuit 620 can apply a signal to the second sub circuit 630 indicating that the activation signal has been received as the input signal is determined as the activation signal.

The second sub circuit 630 can perform a comparison operation between the first cumulative number of times of reception of the activation input signal and the learning threshold value as the signal indicating reception of the activation signal is received from the first sub circuit 620 have.

The second subcircuit 630 may increase the first cumulative number of times of reception each time it receives the activation input signal. Also, the second sub circuit 630 can reset the first cumulative reception count by determining that the first cumulative reception count is equal to or greater than the learning threshold.

Also, the second sub circuit 630 can increase the learning threshold value by determining that the first cumulative reception count is equal to or greater than the learning threshold. According to one embodiment, the learning threshold may have an upper limit of the learning threshold. The second sub circuit 630 may maintain the learning threshold at the upper limit of the learning threshold if the learning threshold is equal to the upper limit of the learning threshold. As another example, the learning threshold may be a fixed value.

The second sub circuit 630 may apply a signal requesting learning to the third sub circuit 640 as it determines that the first accumulation times are equal to or greater than the learning threshold.

The third subcircuit 640 may perform a learning operation to adjust the synapse weight value as it receives the learning request signal from the second subcircuit 630.

The third subcircuit 640 may perform reinforcement learning to set the synaptic weight value corresponding to the active input signal to the probable active value.

For example, the third sub-circuit 640 may obtain information about an input event when an input event that receives an input signal occurs. The information about the input event may include identification information of the synapse pre-neuron circuit included in the input signal. As described above, the identification information of the synapse pre-neuron circuit may include, but is not limited to, synapse information such as cell address information stored with a synapse weight value corresponding to the synapse pre-neuron circuit.

The third subcircuit 640 may perform reinforcement learning to set the synaptic weight value to a probable active value in a synaptic weight memory 610 corresponding to the input signal. For example, the third sub-circuit 640 may store information about a plurality of input events. Also, the third sub-circuit 640 may stochastically determine whether to perform synaptic enhancement that sets the synaptic weight value corresponding to the input event of at least some of the plurality of input events to the active value.

For example, the third subcircuit 640 may independently determine whether to perform synaptic strengthening for each of at least some input events that are to be enforced reinforcement learning.

The first probability that the synapse enrichment probability can be predetermined. For example, the first probability may be a constant, but is not limited thereto.

In addition, the third sub circuit 640 may perform reinforcement learning on a predetermined number of input events among a plurality of previously stored input events. In this case, the third sub-circuit 640 may perform reinforcement learning on a predetermined number of input events in a reverse order of the stored order among the plurality of previously stored input events. If the number of stored input events is less than the predetermined number, the third sub circuit 640 may perform reinforcement learning on all of the previously stored plurality of input events.

The third subcircuit 640 may perform weak learning to set the synaptic weight value to a probable inactive value within the synaptic weight memory 610. [ For example, the third sub-circuit 640 may stochastically determine whether to perform a synapse weakening that sets each of the plurality of synapse weight values in the synapse weight memory 610 to an inactive value.

The third subcircuit 640 may independently determine whether to perform synaptic weakness for each of a plurality of synapse weight values.

A second probability of synapse attenuation probability may be determined based on a plurality of synapse weight values. For example, the second probability may be determined based on the number of activated synapses among which the synapse weight value is the active value.

The third sub-circuit 640 may perform weak learning on at least some of the plurality of synapses in the synaptic weight memory 610. For example, the third sub-circuit 640 may perform weakening learning on all of a plurality of synapses in the synaptic weight memory 610. [ Alternatively, the third sub-circuit 640 may perform weakening learning on synapses in which reinforcement learning is not performed among a plurality of synapses in the synapse weight memory 610.

7 is a block diagram of a neuron circuit 700 according to one embodiment.

The neuron circuit 700 may include a synapse weight memory 710, a first sub circuit 720, a second sub circuit 730, a third sub circuit 740 and a fourth sub circuit 750.

The synapse weight memory 710, the first sub circuit 720, the second sub circuit 730 and the third sub circuit 740 of FIG. 7 are connected to the synapse weight memory 610, the first sub circuit 620, the second subcircuit 630 and the third subcircuit 640 may be applied, and redundant descriptions are omitted.

The first subcircuit 720 may apply a signal to each of the second subcircuit 730 and the fourth subcircuit 750 indicating that the activation input signal has been received.

The second sub circuit 730 may include a learning counter 731, a learning threshold value counter 732, and a first comparison operation unit 733. [

The learning counter 731 may count and output the first cumulative number of times of reception of the activation input signal. For example, the output count value of the learning counter 731 may increase as a signal indicating that the activation input signal is received from the first sub-circuit 720 is received.

The learning threshold value counter 732 may count and output a learning threshold value. The output count value of the learning threshold counter 732 may have a predetermined initial value. The output count value of the learning threshold value counter 732 may be increased based on the output value of the first comparison operation unit 733. [

The first comparison operation unit 733 may receive the output count value of the learning counter 731 and the output count value of the learning threshold value counter 732 to perform a comparison operation. The first comparison operation unit 733 can output a learning request signal when the output count value of the learning counter 731 is equal to the output count value of the learning threshold value counter 732. [ For example, the first comparison operation unit 733 can output a 1-bit signal. 1 is output when the output count value of the learning counter 731 is equal to the output count value of the learning threshold value counter 732, and 0 is output in the other case.

The output value of the first comparison operation unit 733 may be input to the learning threshold counter 732. For example, the output count value of the learning threshold value counter 732 may increase when the output value of the first comparison operation unit 733 is toggled from 0 to 1.

The output value of the first comparison operation unit 733 may be applied to the learning counter 731 as a reset signal. For example, when the output value of the first comparison operation unit 733 is 1, the output count value of the learning counter 731 can be reset to zero.

The third subcircuit 740 may include a random constant generator 741, an enhanced learning processor 742, a weakened learning processor 743, and an input event buffer 744.

The third subcircuit 740 may include a random constant generator 741 that generates a random range of a predetermined range. For example, when the random-number generator 741 outputs a value of 10 bits, the random-number generator 741 may output a random constant ranging from 0 to 1023. For example, the random constant generator 741 may be, but is not limited to, a linear feedback shift register (LFSR).

The input event buffer 744 can store information about an input event when an input event occurs in which the neuron circuit 700 receives an input signal. For example, the input event buffer 744 may store information about a plurality of input events. The information about the input event may include identification information of the synapse pre-neuron circuit included in the input signal. As described above, the identification information of the synapse pre-neuron circuit may include, but is not limited to, synapse information such as cell address information stored with a synapse weight value corresponding to the synapse pre-neuron circuit. When the input event buffer 744 is full, the input event buffer 744 may delete the input event information having the fastest storage order and store information on a new input event. For example, the input event buffer 744 may be a circular buffer.

As the third subcircuit 740 receives the training request signal from the second subcircuit 730, reinforcement learning may be performed using the reinforcement learning processor 742. [

The reinforcement learning processor 742 may perform reinforcement learning to set the synaptic weight value corresponding to the previously received input signal as a probabilistic active value. For example, reinforcement learning processor 742 may perform reinforcement learning to set the synaptic weight value corresponding to a plurality of input signals that have been received to a probable active value. For example, reinforcement learning processor 742 may determine whether to perform synaptic strengthening to set a synaptic weight value corresponding to information of each of at least some of the plurality of input events stored in input event buffer 744 to an active value Can be determined stochastically.

For example, the reinforcement learning processor 742 may independently determine whether to perform synaptic enhancement for each of at least some of the input events that are to be enforced reinforcement learning.

The first probability that the synapse enrichment probability can be predetermined. For example, the first probability may be a constant, but is not limited thereto.

For example, reinforcement learning processor 742 may perform reinforcement learning on a predetermined number of input events of a plurality of input events previously stored in input event buffer 744. [ At this time, the reinforcement learning processor 742 can perform reinforcement learning on a predetermined number of input events in a reverse order of the stored order among the plurality of input events previously stored in the input event buffer 744. [ If the number of input events pre-stored in the input event buffer 744 is less than the predetermined number, the reinforcement learning processor 742 may perform reinforcement learning on all of the pre-stored input events in the input event buffer 744 .

The enhanced learning processor 742 may determine whether to perform synaptic enhancement by comparing the random constant obtained from the random constant generator 741 with a reinforcement learning reference constant determined based on the first probability. For example, the enhanced learning processor 742 may perform synaptic enhancement if the random constants are less than or equal to the enhanced learning criterion constant calculated by multiplying the upper limit value of the random constants by the first probability. That is, the reinforcement learning processor 742 may set the synaptic weight value of the synaptic weight memory 710, corresponding to learning target input event information, to the active value. The reinforcement learning processor 742 may apply the cell address information, the write request signal, and the activation value corresponding to the input event to the synapse weight memory 710. [

As another example, the enhanced learning processor 742 may perform synaptic enhancement if the random constant is greater than or equal to or greater than the enhanced learning criterion constant.

The weakened learning processor 743 may perform weak learning to set the synaptic weight value in the synapse weight memory 710 to a stochastically inactive value. The weakened learning processor 743 may perform weakened learning after the enhanced learning is performed by the enhanced learning processor 742. [

The weakened learning processor 743 may perform weak learning on the synapses of at least some of the plurality of synapses in the synaptic weight memory 710. [

For example, the weakened learning processor 743 may perform weak learning on all of the plurality of synapses in the synaptic weight memory 710. Alternatively, the weakened learning processor 743 may perform weak learning on synapses for which reinforcement learning has not been performed among a plurality of synapses in the synaptic weight memory 710.

The weakened learning processor 743 may stochastically determine whether to perform a synapse weakening that sets each of the plurality of synapse weight values in the synapse weight memory 710 to an inactive value.

For example, the weakened learning processor 743 may independently determine whether to perform a synapse weakness for a plurality of synapse weight values.

The weakened learning processor 743 may determine a second probability that is a synapse weakening probability based on the plurality of synapse weight values. For example, the second probability may be determined based on the number of activated synapses among which the synapse weight value is the active value. For example, the second probability may be a ratio of the difference between the number of activated synapses to the number of activated synapses to a predetermined number of activated synapses.

The weakened learning processor 743 may determine whether to perform the synapse weakness by comparing the random constant obtained from the random constant generator 741 with the weakened learning reference constant determined based on the second probability. For example, the weakened learning processor 743 may perform a synapse weakening if the random constant is less than the weakened learning criterion constant calculated by multiplying the upper limit value of the random constant by the second probability. That is, if the random constant is less than or equal to the weakened learning reference constant, the weakened learning processor 743 may set the synapse weight value of the synapse weight memory 710 to an inactive value. The weakening learning processor 743 may apply the cell address information, the write request signal and the inactivity value corresponding to the synaptic weight value to be weakened to the synapse weight memory 710. [

As another example, the weakened learning processor 743 may perform a synapse weakness if the random constant is greater than, equal to, or greater than the weakening criterion constant.

The fourth sub-circuit 750 may perform a comparison operation between the second cumulative number of times of reception of the activation input signal and the ignition threshold value. In addition, the fourth sub-circuit 750 may transmit the spike signal 71 to the post-synaptic neuron circuit of the neuron circuit 700 when the second cumulative number of times of reception of the activation input signal reaches the utterance threshold value.

The fourth sub circuit 750 may include an ignition counter 751 and a second comparison operation unit 752. The ignition counter 751 counts and outputs the second cumulative number of times of reception of the activation input signal. For example, the output count value of the ignition counter 751 may increase as it receives a signal from the first sub-circuit 720 indicating that the activation input signal has been received.

The second comparison operation unit 752 may perform a comparison operation by receiving the output count value of the utterance counter 751 and the utterance threshold value. The ignition threshold may be a predetermined constant. For example, the fire threshold may be a pre-stored value or a value received from an external component.

The second comparison operation unit 752 can output the spike signal 71 when the output count value of the utterance counter 751 is equal to the utterance threshold value. For example, the second comparison operation unit 752 can output a 1-bit signal. At this time, the second comparison operation unit 752 outputs 1 when the output count value of the ignition counter 751 is equal to the ignition threshold value, and outputs 0 when the output count value is equal to the ignition threshold value. When the output value of the second comparison operation unit 752 is 1, it indicates that the spike signal 71 is outputted.

The output value of the second comparison operation unit 752 may be applied to the ignition counter 751 as a reset signal. For example, when the output value of the second comparison operation unit 752 is 1, the output count value of the ignition counter 751 can be reset to zero.

The spike signal 71 of the second comparison operation unit 752 may be applied to the neuron circuit after the synapse of the neuron circuit 700. [

8 is a diagram illustrating a synapse weight learning system 800 in accordance with one embodiment.

The synapse weight learning system 800 may include a neuron circuit 860 and a learning circuit 840.

8, the learning circuit 840 outside the neuron circuit 860 performs a synapse weight learning operation in the neuron circuit 700. The learning circuit 840, which is outside the neuron circuit 860, Weight learning operation can be performed.

The synaptic weight memory 610 and 710 and the first subcircuits 620 and 720 of FIGS. 6 and 7 may be applied to the synapse weight memory 810 and the first subcircuit 820 of FIG. 8, respectively . The second subcircuit 630 of FIG. 6 and FIG. 7 is connected to the second subcircuit 830, the learning counter 831, the learning threshold value counter 832, and the first comparison operation unit 833 of FIG. 8, , 730, a learning counter 731, a learning threshold value counter 732, and a first comparison operation unit 733 can be applied. Each of the fourth sub-circuit 850, the utterance counter 851 and the second comparison operation section 852 of Fig. 8 includes the fourth sub-circuit 750, the utterance counter 751, (752) may be applied.

The learning circuit 840, the random-number generator 841, the reinforcement learning processor 842, the weakening-learning processor 843, and the input event buffer 844 shown in Fig. Embodiments of the random variable generator 640 and 740, the random constant generator 741, the reinforcement learning processor 742, the weakening learning processor 743, and the input event buffer 744 can be applied. Therefore, in the description of FIG. 8, the description overlapping with FIG. 6 and FIG. 7 will be omitted.

The neuron circuit 860 determines whether the received input signal is an activation signal whose active value is the synapse value of the received input signal. If the first accumulated reception number of the activated input signal reaches the learning threshold value, To the circuit.

The neuron circuit 860 can transmit the learning request signal of the first comparison operation unit 833 included in the second sub circuit 830 to the learning circuit 840. [ Further, the neuron circuit 860 can transmit the identification information of the neuron circuit 860 to the learning circuit 840 together with the learning request signal. Thus, the learning circuit 840 can identify the neuron circuit 860 that transmitted the learning request signal based on the identification information of the neuron circuit 860. [

As the learning circuit 840 receives the learning request signal from the neuron circuit 860, each of the reinforcement learning processor 842 and the weakened learning processor 843 accesses the synapse weight memory 810 of the neuron circuit 860 , Reinforcement learning operation, and weakening learning operation.

9 is a diagram illustrating a synapse weight learning system for a plurality of neuron circuits according to one embodiment.

An embodiment for the neuron circuit of Fig. 8 may be applied to each of the plurality of neuron circuits, and an embodiment of the learning circuit of Fig. 8 may be applied to the learning circuit.

Learning circuit 940 may perform a synapse weight learning operation on a plurality of neuron circuits 910 and 920. Further, each of the plurality of neuron circuits 910 and 920 can be mutually identified through the identification information.

A plurality of neuron circuits 910 and 920 and a learning circuit 940 can communicate with each other via a bus (BUS) Learning circuit 940 may receive a learning request signal from each of a plurality of neuron circuits 910 and 920 via bus 970. Upon receiving the learning request signal and the identification information of the neuron circuit from the neuron circuits 910 and 920, the learning circuit 940 can transmit and receive information to and from the neuron circuit corresponding to the identification information using the identification information. Accordingly, the learning circuit 940 can perform the synapse weight learning operation of the neuron circuit corresponding to the identification information via the bus.

10 is a diagram showing a learning system of a hierarchical neuron circuit group.

The learning system of a neuronal circuit population may include a plurality of hierarchical groups. Each of the hierarchical groups may comprise a plurality of neuron circuits and a learning circuit.

The spike signals output from each of the neuron circuits of the one-tier cluster may be applied as input signals to the next tier cluster.

For example, a spike signal (spike signal N-1_1, spike signal N-1_1) output from each of the neuron circuits (neuron circuit N-1_1, neuron circuit N-1_2, N-1_2, ..., spike signal N-1_A) may be applied to the input signal (input signal N) of the N-tier cluster.

Likewise, the spike signals (spike signal N_1, spike signal N_2, ..., spike signal N_A) output from each of the neuron circuits (neuron circuit N_1, neuron circuit N_2, (Input signal N + 1) of the +1 hierarchical group.

Also, the learning circuit of each hierarchical group can perform a synapse learning operation on a plurality of neuron circuits of the hierarchical group. For example, the learning circuit N-1, the learning circuit N, and the learning circuit N + 1, respectively, can perform the synapse learning operation of the neuron circuits belonging to each of the N-1 hierarchical group, the N-hierarchical group, and the N + . Multiple neuron circuits of FIG. 10 Each of the learning circuits may be applied to the neuron circuit and learning circuit of FIGS. 8 and 9, respectively.

11 is a flow diagram of a method of learning a synaptic weight performed by a neuron circuit according to one embodiment.

In step 1110, the neuron circuit may receive an input signal from the synaptic neuron circuit and determine whether the received input signal is an activation signal whose synaptic weight value of the received input signal is the active value.

The neuron circuit may store a plurality of synapse weight values. For example, a neuron circuit may store a synaptic weight value corresponding to each of a plurality of synapse pre-neuron circuits. For example, the synaptic weight value may be one bit of information. The synapse weight value may have an active value or a binary value of an inactive value. For example, the active value may be 1 and the inactive value may be 0, but is not limited thereto. At this time, the synapse whose synaptic weight value is the active value is defined as the activated synapse, and the inactive value is defined as the inactivated synapse.

The neuron circuit can receive the input signal. The input signal may be a signal output from the synaptic neuron circuit of the neuron circuit. For example, the input signal may be a spike signal output from a synaptic neuron circuit. In addition, the input signal may be a signal output from a plurality of synaptic pre-neuron circuits. For example, the input signal may be a signal to which spike signals output from at least some of the plurality of synaptic pre-neuron circuits are input selectively or in combination.

For example, the input signal may include information that can identify a synaptic neuronal circuit that has output a spike signal. As another example, the input signal may include information indicating that the synapse pre-neuron circuit has output a spike signal and / or identification information of the synapse pre-neuron circuit. For example, the identification information contained in the input signal may include identification information unique to the synaptic neuron circuit. In addition, the identification information included in the input signal may include cell address information of a memory in which a synapse weight value of the synapse pre-neuron circuit is stored.

The neuron circuit can determine whether the input signal is an active input signal whose synaptic weight value corresponding to the input signal is the active value. For example, the neuron circuit may obtain a synapse weight value corresponding to the input signal. If the obtained synapse weight value is the active value, the neuron circuit can determine that the input signal is an activation signal. If the obtained synapse weight value is an inactive value, the neuron circuit can determine that the input signal is not an activation signal.

In step 1120, as the input signal is determined to be an activation signal, the neuron circuit may perform a comparison operation between the first cumulative number of times of reception of the activation input signal and the learning threshold.

The neuron circuit can reset the first cumulative number of times of reception according to the determination that the first cumulative number of receptions has reached the learning threshold value.

For example, the neuron circuit may increase the learning threshold as it determines that the first cumulative number of receptions has reached the learning threshold. According to one embodiment, the learning threshold may be the upper limit of the learning threshold. The neuron circuit can keep the learning threshold at the upper limit of the learning threshold if the learning threshold is equal to the upper limit of the learning threshold. As another example, the learning threshold may be a fixed value.

In step 1130, the neuron circuit may perform reinforcement learning to set the synaptic weight value of the at least one input signal received to the active value, according to the first probability, if the first cumulative number of receptions reaches a learning threshold have.

For example, as determining that the first cumulative number of received times has reached the learning threshold, the neuron circuit may perform reinforcement learning to adjust the synapse weight value.

The neuron circuit may perform reinforcement learning to set the synaptic weight value corresponding to the received input signal to the active value stochastically.

The neuron circuit can obtain information about an input event when an input event for receiving an input signal occurs. The neuron circuit may store information about a plurality of input events. In addition, the neuron circuit may stochastically determine whether to perform synaptic enhancement that sets the synaptic weight value corresponding to the input event of at least some of the plurality of input events to an active value. At this time, the neuron circuit can independently determine whether or not to perform synaptic strengthening for each of at least some input events to be subjected to reinforcement learning.

The information about the input event may include identification information of the synapse pre-neuron circuit included in the input signal. As described above, the identification information of the synapse pre-neuron circuit may include, but is not limited to, synapse information, such as address information, in which the synapse weight value corresponding to the synapse pre-neuron circuit is stored.

The neuron circuit may perform reinforcement learning on a predetermined number of input events among a plurality of pre-stored input events. At this time, the neuron circuit can perform reinforcement learning on a predetermined number of input events in the reverse order of the stored order among the plurality of pre-stored input events. If the number of pre-stored plurality of input events is less than a predetermined number, the neuron circuit may perform reinforcement learning for all of the pre-stored plurality of input events.

The first probability that the synapse enrichment rate can be predetermined. For example, the first probability may be a constant, but is not limited thereto.

The neuron circuit may obtain a random constant, and may determine whether to perform synaptic enhancement by comparing the determined random constant with the determined enhanced learning reference constant based on the first probability. For example, the neuron circuit may perform synaptic enhancement if the random constants are less than the reinforcement learning criterion constant calculated by multiplying the upper limit of the random constants by the first probability. At this time, the random constant may be a positive number, and is a random number determined among positive numbers in a predetermined range.

As another example, the neuronal circuit can perform synaptic enhancement if the random constant is less than the enhanced learning criterion constant.

In step 1140, the neuron circuit may perform weak learning to set each of the plurality of synapse weight values to an inactive value, according to a second probability.

The neuron circuit can perform weakening learning to set the synapse weight value to a stochastic inactive value. For example, the neuron circuit may stochastically determine whether to perform a synapse weakening that sets each of the plurality of synapse weight values to an inactive value. At this time, the neuron circuit can independently determine whether synaptic weakening is to be performed for each of a plurality of synapse weight values.

The neuron circuit can perform weak learning after performing reinforcement learning. The neuron circuit may perform weakening learning on at least some of the plurality of synapses. For example, a neuron circuit can perform weakening learning for all of a plurality of synapses. As another example, the neuron circuit may perform weakening learning on a synapse where reinforcement learning is not performed among a plurality of synapses.

For example, the neuron circuit may stochastically determine whether to perform a synapse weakening that sets each of the weight values of a plurality of synapses to an inactive value.

The neuron circuit may determine a second probability that is a synapse attenuation probability based on the plurality of synapse weight values. For example, the second probability may be determined based on the number of activated synapses among which the synapse weight value is the active value. For example, the second probability may be a ratio of the number of predetermined activated activation synapses to the number of activated synapses to the number of activated synapses.

The neuron circuit can obtain a random constant, and compare the determined random constant with the determined weakened learning reference constant based on the second probability to determine whether to perform a synapse weakening. For example, the neuron circuit may perform synaptic weakness if the random constants are less than the weakened learning criterion constant calculated by multiplying the upper limit of the random constants by the second probability. That is, when the random constant is smaller than the weakened learning reference constant, the neuron circuit can set the weight value of the synapse to the inactive value.

As another example, the neuron circuit can perform synaptic weakness when the random constant is greater than the attenuation criterion constant.

As the input signal is determined to be an activation signal, the neuron circuit may perform a comparison operation between the second cumulative number of times of reception of the activation input signal and the ignition threshold value. Further, the neuron circuit can output a spike signal when the second cumulative reception count reaches the ignition threshold value. For example, a neuron circuit can transmit a spike signal to a post-synaptic neuron circuit.

The neuron circuit can reset the second cumulative reception count as it is determined that the second cumulative reception count has reached the ignition threshold value.

For example, the ignition threshold may be a fixed value. As another example, the neuron circuit may increase the utterance threshold value as it determines that the second cumulative reception count has reached the utterance threshold. Further, the ignition threshold value may have an upper limit value of the ignition threshold value. The neuron circuit may maintain the ignition threshold at the upper limit of the ignition threshold if the ignition threshold is equal to the upper limit of the ignition threshold.

12 is a diagram showing a scheme of a neuron block for implementing a neuron circuit.

The neuron block can have unique neuron address information. For example, the neuron address signal Neuron_addr may include the identification address information of each corresponding neuron block. The STDP learning process for the neuron block is activated when the STDP active neuron address signal STDP_active_addr value, which is externally applied to transmit / receive information to / from the neuron block, coincides with the neuron address information signal Neuron_addr value.

The neuron block can receive the input signal AERin_v. AERin_v can represent a spike signal applied from the synaptic circuit.

Also, the neuron block can receive the input signal AERin. The input signal AERin may include information about the synapse corresponding to the entire circuit for outputting the received input signal AErin_v. For example, AERin may represent address information of an input terminal receiving a corresponding input signal among a plurality of input terminals applying an input signal to a neuron block. Alternatively, AERin may represent address information of a synapse circuit that applies an input signal to a neuron block. Alternatively, the AERin signal may include address information of a Synaptic Weight Memory cell that stores a corresponding synapse weight value.

Also, the neuron activation signal Neuron_active may indicate that the neuron block is in a steady state capable of processing the input signal. For example, if the neuron activation signal Neuron_active has a value of 0, even if the neuron block receives an input from the pre-synapse circuit, the counting trigger signal Count-up1, which is input to the Firing Counter and the Learning Counter, The output count value of the Firing Counter and the Learning Counter may not increase.

When the value of the neuron activation signal Neuron_active is 1, the output count value of the Firing Counter and the Learning Counter can be increased as the input signal AERin_v having the active value, for example, have. For example, as the input signal AERin_v having the active value is input, the value of the counting trigger signal Count-up1 input to the Firing Counter and the Learning Counter can be toggled to 1, The output count value of the Firing Counter and the Learning Counter may increase.

At this time, the synapse weight value corresponding to the input signal AERin_v can be considered so that the counting trigger signal Count-up1 value is toggled to 1. For example, the neuron block may obtain a corresponding synapse weight value syn_weight from the synapse weight memory based on the synaptic address signal rd_addr contained in the AERin signal corresponding to the input signal AERin_v. The synapse address signal rd_addr is used to acquire a synapse weight value syn_weight used in the neuron block for processing of the input signal AERin_v and distinguishes it from the STDP request synapse address signal STDP_RD_addr used for information transmission and reception with the STDP learning block .

The count trigger signal (first count trigger signal) Count_up1 value can be toggled to 1 only when the synaptic weight value syn_weight obtained according to the input signals AERin_v and AERin_v is 1. This indicates that the synapse transmitting that input is active.

On the other hand, the neuron block includes a Synaptic Weight Memory. For example, the synaptic weight memory has a capacity of 1024 bits, and thus can store synapse weight values corresponding to a maximum of 1024 synapses. For example, the synaptic weight value stored in each synapse weighted memory cell may have a value of either an active value or an inactive value. For example, the active value of the synapse weight value may be one, and the inactive value may be zero. At this time, the initial values of all synapse weights stored in the synaptic weight memory may be set to one.

The Firing Counter can receive the counting trigger signal Count-up1 as the input signal, and the reset signal reset which initializes the count value. The reset signal reset may have a value for resetting the count value of the firing counter when the spike signal Spike_out is output or when receiving a signal inh_event for commanding initialization from the outside.

The count value count2 of the Firing Counter is used as a criterion for determining whether or not the neuron is ignited. For example, if the count value count2 of the Firing Counter reaches the predetermined spike threshold spike_threshold, the neuron can be ignited. In this case, the ignition threshold value spike_threshold may be a predetermined fixed value. As another example, the spark threshold value spike_threshold may increase every time the count value count2 of the Firing Counter reaches the spike threshold value. At this time, even if the spiking threshold value spike_threshold increases, there is no upper limit value or there may be a predetermined upper limit value.

For example, a comparison operation between the count value count2 of the Firing Counter and the spike-threshold can be performed. The spike signal Spike_out value is 1 when the count value count2 of the Firing Counter is equal to or greater than the spike threshold value. A spike signal Spike_out value of 1 indicates that the neuron block outputs a spike signal. For example, the output spike Spike_out signal can be transmitted to the neuron block of the upper layer. For example, in a hierarchical neuron system, a spike signal Spike_out output from a current neuron block is transmitted to a neuron block of an upper layer, which is a neuron block that is a current neuron block, Lt; / RTI &gt;

In addition, the neuron module may include a learning request module that requests learning to the STDP learning module. For example, the neuron block may include a Learning Counter and a STDP Threshold Counter (STDP Threshold Counter) for determining whether to request a learning. Since the learning rate of the current neuron block is gradually decreased as the ignition counter and the learning counter used for the ignition of the neuron block of the upper layer or the learning request are distinguished from each other, Can be prevented from being lowered.

Like the ignition counter, the learning counter can increase the count value when the first count trigger signal count-up1 toggles to one.

When the count value count1 of the learning counter reaches a predetermined learning threshold value, the neuron block can output a learning request signal STDP_req for requesting learning. For example, a comparison operation between the count value count3 of the STDP threshold counter (STDP_threshold counter) and the count value count1 of the learning counter (Learning Counter), which outputs the learning threshold described below, can be performed. As a result of the comparison operation, if the count value count1 of the learning counter is equal to or larger than the count value count3 of the STDP threshold counter (STDP threshold counter), a learning request STDP_req signal having a value of 1 can be output. For example, a STDP_req value of 1 indicates that a learning request is made.

The learning counter may receive a reset signal. For example, a reset signal value may be determined based on signals STDP_activate, STDP_event indicating whether STDP learning has been activated or performed and STDP_req signal indicating whether the neuron block has requested STDP learning. For example, when the counting value count1 of the learning counter reaches the count value count3 of the STDP threshold counter (STDP_threshold counter) and the STDP_req signal value indicating the learning request becomes 1, the learning counter is reset And the count value count1 may be reset to an initial value, for example, zero.

The learning threshold can be increased. For example, the learning threshold value may be the count value count3 output from the STDP threshold counter (STDP_threshold Counter).

The STDP threshold counter (STDP_threshold Counter) can acquire an initialization signal initializ as an input. For example, the initialization signal initialize may be a predetermined initial learning threshold STDP_threshold_init. The STDP threshold counter (STDP_threshold Counter) can output the initial learning threshold value STDP_threshold_init as the count value count as it receives the initialization signal initialize.

Also, the STDP threshold counter (STDP_threshold Counter) can receive the second count trigger signal count_up2. The count value count3 of the STDP threshold counter (STDP_threshold Counter) may increase as the second count trigger signal count_up2 toggles to one.

The second count trigger signal can be determined based on the learning request signal STDP_req, which is a signal for requesting learning in the STDP block. As the learning request signal STDP_req is toggled to 1, the Count_up2 signal having the same value as the STDP_req signal is toggled as 1, and the count value count3 of the STDP threshold counter (STDP_threshold counter) can be increased.

The increase in the count value count3 of the STDP threshold counter (STDP_threshold counter) may mean that the learning rate of the neuron gradually decreases as learning progresses. This prevents the concentration of learning of excessive features on specific neurons and causes competition between neurons to occur. As another example, the learning threshold may be a predetermined fixed value. As described above, since a firing counter serving as an output reference of the spike signal spike_out used for the learning of the neuron block of the upper layer exists separately, the learning rate of the neuron block of the upper layer is determined by the learning threshold of the current block Is not directly affected by the learning speed degradation of the current neuron block that occurs as the number of neurons increases.

In addition, the neuron block can perform a leak process.

The leak event signal Leak_event can be input to the Firing Counter and the Learning Counter. Leak event signal Leak_event is to implement state change of neuron over time. For example, the counting values count1 and count2 of the Firing Counter and the Learning Counter may be decreased at predetermined time intervals according to the leakage event signal Leak_event.

Likewise, as the threshold leak event signal TH_Leak_event is input, the count value of the STDP threshold counter (STDP_threshold Counter) may decrease every predetermined time.

At this time, the decreasing rate of the count value of the STDP threshold counter (STDP threshold counter) due to the threshold leak event signal TH_Leak_event is smaller than the decreasing rate of the counter value (Firing Counter) caused by the leak event signal Leak_event and the count value of the learning counter Can be slower.

As described above, the STDP_active_addr signal, the STDP_RD_addr signal, the STDP_WR_en signal, and the STDP_WR_data signal are input signals received from the STDP learning block.

For example, the input signals described above can be used to allow the STDP learning block to access the Synaptic Weight Memory within the neuron block and read or change the weight value.

The STDP activated neuron address signal STDP_active_addr is an input signal received from the STDP learning block and contains address information of the STDP learning target neuron. When the values of the two signals are equal to each other through a comparison operation between the value of the STDP activated neuron address signal STDP_active_addr and the value of the neuron address information signal Neuron_addr including the unique address information of the neuron block, the STDP learning block stores the synapse weight memory Can be met.

The STDP learning block needs to access each of the plurality of weight values stored in the synaptic weight memory to perform the learning process. At this time, the STDP request synapse address signal STDP_RD_addr indicates cell address information for accessing the corresponding cell in order to read or change the synapse weight value stored in the specific cell by the STDP learning block. The term synapse address information includes cell address information.

The synapse weight memory STDP_RD_addr can be received from the STDP learning block. When the write enable signal STDP_WR_en has an inactive value, for example, 0, the synapse weight memory outputs the weight value of the synapse stored in the cell corresponding to the STDP request synapse address signal STDP_RD_addr value as STDP_RD_data. The output synaptic weight value is transmitted to the STDP learning block through the STDP read signal STDP_RD_data and used for learning.

The STDP_WR_addr input to the synapse weight memory is a write synapse address signal and is determined based on the STDP request synapse address signal STDP_RD_addr. At this time, the write synapse address signal STDP_WR_addr has a valid address value only when the STDP write enable signal STDP_WR_en received from the STDP learning block is 1 to activate the write function to the synapse weight memory.

When the write enable signal STDP_WR_en has an active value, the weight value of the synapse stored in the cell corresponding to the write synapse address signal STDP_WR_addr value is set (changed) to the inputted STDP write signal STDP_WR_data value.

Using the above-described read and write process, the STDP learning block can access the neuron block to be learned, access the synaptic weight memory within the neuron block to be learned, read the synapse weight value, You can set the value.

13 is a diagram illustrating an STDP learning block scheme for implementing a learning circuit according to an embodiment.

Referring to FIG. 13, the STDP learning block may include a circular buffer for storing input events.

In the STDP learning block, the STDP learning block may include an enhanced learning processor (LTP processor) for performing long-term reinforcement learning, and a weakened learning processor (LTD Processor) for performing weakening learning.

In addition, the STDP learning block may include a main processor that manages the start of the learning process according to the request of the STDP learning, the start and end of the reinforcement learning, and the start and end of the weakened learning.

As at least one neuron block of the plurality of neuron blocks transmits the learning request signal, the STDP learning block can receive the learning request reception signal STDP_addr_v having the active value. Also, it can receive the learning request neuron address signal STDP_req_addr indicating the address of the neuron block that requested the learning. For example, the learning request neuron address signal STDP_req_addr can identify up to 256 neuron blocks as 8-bit signals. That is, one STDP learning block can perform a learning process on 256 neuron blocks.

As the learning request reception signal STDP_addr_v has an active value and receives the ready signal core_rdy signal indicating that the preparation for starting the learning processor from the main processor is completed as described below, The STDP active neuron address signal STDP_active_addr indicating the address information of the learning target neuron block may have a value corresponding to the learning request neuron address signal STDP_req_addr.

Then, the STDP learning block can access the neuron block corresponding to the STDP activated neuron address signal STDP_active_addr, and performs a learning process on the neuron block.

As the learning process is started, the main processor can apply the reinforcement learning start signal to the reinforcement learning processor (LTP Processor). The reinforcement learning processor (LTP processor) then performs the reinforcement learning, and when the reinforcement learning is completed, the reinforcement learning completion signal end can be applied to the main processor.

The main processor receiving the reinforcement learning completion signal end from the reinforcement learning processor (LTP processor) can sequentially apply the weakening learning start signal to the weakening learning processor (LTD Processor). Then, the weakened learning processor (LTD Processor) performs the weakened learning and when the weakened learning is completed, the weakened learning completion signal end can be applied to the main processor.

Upon receipt of the weakened learning completion signal, the main processor can complete learning of the neuron block.

While the enhanced learning processor (LTP Processor) performs the reinforcement learning, the main processor is configured such that the output value associated with reinforcement learning, among the output values associated with reinforcement learning and output values associated with the weakening learning, As shown in FIG. For example, in order for the multiplexer (LTP / LTD), which selectively outputs output values related to reinforcement learning and output values associated with weakening learning, to output output values related to reinforcement learning, the reinforcement learning activation signal LTP_active And can be applied to the multiplexer LTP / LTD. For example, if the reinforcement learning activation signal LTP_active is an active value, the multiplexer (LTP / LTD) outputs an output value associated with reinforcement learning and outputs an output value associated with the weakening learning if the reinforcement learning activation signal LTP_active is an inactive value.

Upon receiving the reinforcement learning start signal start from the main processor, the reinforcement learning processor starts reinforcement learning.

The reinforcement learning is performed based on the data stored in the circular buffer storing the input event information. The pre-event information may indicate, for example, synapse address information or input terminal address information corresponding to input signals applied to a learning target block.

The circular buffer sequentially stores an AERin signal including corresponding synapse address information each time it receives an input signal AERin_v applied to at least one neuron block including learning target neurons. The circular buffer can have a total capacity of 1024x10 bits. That is, the circular buffer can store 10 bits of synapse address information corresponding to 1024 input signals that have recently occurred. At this time, it may be an address capable of identifying a plurality of input signal terminals transmitting an input signal to the synapse address information neuron block. Or the synapse address information may be the address information of the synapse preneuron block that transmitted the spike signal, which is the input signal. The combination of the address information of the input terminal or the address information of the synapse precursor block and the address information of the target neuron block receiving the input signal may represent the synapse address information. As the active value is received as AERin_v, the circular buffer write enable signal wr_en applied to the circular buffer may have an active value.

The enhanced learning processor may set the synapse weight value corresponding to the recently stored synapse address information among the synapse address information stored in the circular buffer to an active value based on the predetermined enhancement probability determination factor LTP probability.

At this time, the reinforcement learning processor can perform reinforcement learning using a predetermined number of input events occurring recently among input events corresponding to synapse address information stored in the circular buffer. The predetermined number may be the number of reinforcement learning number of the input value. The reinforcement learning processor may increase the reinforcement learning counter count value by 1 (LTP_cnt + 1) by applying the reinforcement learning count trigger signal cnt_up value every time reinforcement learning is performed based on each input event stored in the circular buffer. If a comparison operation is performed between the increased reinforcement learning count value and the reinforcement learning number number of potentiation and is the same, the reinforcement learning processor can receive the reinforcement learning completion signal LTP_done having a value indicating completion. The enhanced learning completion signal LTP_done may have a value indicating completion even by the signal CB_empty indicating that there is no pre-event stored in the circular buffer output from the circular buffer. That is, the reinforcement learning processor ends the reinforcement learning when the circular buffer is empty or when the number of pre-events for which reinforcement learning has been performed reaches the number of reinforcement learning number of potencies.

The enhanced learning processor may apply a circular buffer read enable signal CB_rd_en to the circular buffer. Accordingly, the circular buffer can output the synaptic address information corresponding to the input event of the current sequence to the circular buffer output signal CB_dout. The circular buffer output signal CB_dout is output as an intermediate signal rd_addr through a multiplexer (LTP / LTD).

Further, the reinforcement learning processor can output the reinforcement learning write enable signal LTP_wr_en. Since the reinforcement learning is in progress, the intermediate signal rd_addr for reinforcement learning and the reinforcement learning write enable signal LTP_wr_en signal can be output as an output signal of the LTP / LTD multiplexer in accordance with LTP_active. Depending on the output value of the far-away plexer, the STDP learning block may output the intermediate signal rd_addr as the STDP request synapse address signal STDP_RD_addr and output the reinforcement learning write enable signal LTP_wr_en as the write enable signal STDP_WR_en.

Accordingly, the synapse weight value stored in the cell in the synapse weight memory corresponding to the STDP request synapse address signal STDP_RD_addr from the learning target neuron block can be obtained through the STDP_RD_data signal. During the read process, the enhanced learning processor can output the write enable signal STDP_WR_en value as an inactive value. In other words, during the read process, the enhanced learning processor can disable the write enable signal STDP_WR_en.

Also, a comparison operation between the random constant generated by the random constant generator and the predetermined enhancement probability determining factor LTP probability is performed so that the synapse weight can be enhanced if the generated random constant is equal to or less than the enhancement probability determining factor LTP probability. Indicates that the synapse weight value corresponding to the STDP request synapse address signal STDP_RD_Addr, which intensifies the synapse weight, is set to one. That is, the STDP learning block outputs 1 as the STDP write data signal STDP_WR_data value to the learning target neuron. The STDP write data signal STDP_WR_data value is transmitted to the learning target block, and the synapse weight value stored in the cell corresponding to the synapse weight memory of the learning target block object is set to one. During the write process, the reinforcement learning processor may output 1 as the write enable signal LTP_wr_en. Accordingly, the STDP learning block can output 1 as the STDP write enable signal STDP_WR_en.

After reinforcement learning, the weakened learning processor may perform weak learning on the target neuron block.

The weakened learning processor may sequentially perform the weakened learning process for each of all the cells storing the synaptic weight value in the synapse weight memory within the learning target neuron block. Thus, the weakened learning processor sequentially increases the output value of the rd_counter used for determining the address value of the cell, in order to sequentially access each cell in the synapse weight memory. Accordingly, the STDP request synapse address signal STDP_RD_addr is sequentially changed and output in accordance with the sequentially increased rd_counter output value. Thus, the STDP learning block can access each of all the cells in the synaptic weight memory of the learning neuron.

First, the STDP learning block sequentially receives each of the synapse weight values stored in all the cells of the synapse weight memory in the learning target neuron block, and applies the input weight value to the sum signal increase_cnt_up. Therefore, the output value of the weight sum counter (Cnt_Sum of weight) may be the total sum of the synapse weight values stored in the synapse weight memory.

Also, the STDP learning block can receive a predetermined number of predictive synapse weights, Number of active. Also, the STDP learning block calculates the difference between the total sum of the synaptic weight values, i.e., the output value of the weight sum counter (Cnt_Sum of weight) in the predictive synapse weight sum Number of Active, multiplies the calculated value by 2 ^ 10, This can be calculated by dividing by the total sum of the synaptic weight values, the degradation probability determination factor LTD Probability.

Using the calculated weakness probability factor LTD Probability, the weakened learning processor can perform stochastic weakening learning for each cell of the synaptic weight memory within the learning target neuron block.

The weakening learning processor sequentially outputs the address information corresponding to each cell of the synapse weight memory as STDP_RD_addr to acquire the synapse weight value of the corresponding cell, and based on the calculated weakening probability determination factor LTD Probability, Can be changed to zero. Specifically, a comparison operation between the random constant generated by the random constant generator and the weakening probability determination factor LTD Probability is performed. If the generated random constant is greater than the weakening probability determination factor LTD Probability, the weight value of the corresponding cell is changed to zero.

For example, when the generated random constant is larger than the weakening probability determining factor LTD Probability, 0 is applied to the STDP write data signal STDP_WR_data, and the weight value of the cell corresponding to the current STDP request synapse address signal STDP_RD_addr can be set to 0 . The weakening learning described above is performed for each of all the cells in the synapse weight memory. However, the random constant may be a value generated separately for each of the cells, so that the weakening learning may be performed between the cells in the same Synaptic Memory.

14 is a diagram illustrating a single core block scheme for implementing a single core circuit including a plurality of neuron circuits according to one embodiment.

15 is a diagram showing a multicore block including a plurality of single core blocks.

A single core block may include a plurality of neuron blocks (Neuron0, Neuron1 Neuron2, Neuron255) and a single STDP learning block. In FIG. 14, a single core block is shown to include 255 neuron blocks, but the number of neuron blocks is not limited thereto.

A single core block can receive the AER_in signal. The AER_in signal may be, for example, a signal received from a synaptic neuron block of at least one neuron block contained in a single core block. The AER_in signal may, for example, represent a spike signal output from a synaptic neuron block of at least one neuron block.

The AER filter is capable of filtering signals that are not related to neuron blocks within a single core block of the AERin signal. The AER filter can be implemented according to the architecture of the artificial neural network.

The filtered input signal AERin may be input to at least one neuron block of a plurality of neurons included in a single core block. The AERin signal may be the same as, for example, the input signal 70 of FIG. 7, and the operation described in FIG. 12 may be applied to the operation performed on each neuron block.

In a single core circuit, a single STDP learning block performs a learning process for all of a plurality of neurons, so a method for preventing signal collision between a plurality of neurons is required.

On the other hand, the inputs shared by a plurality of neurons in a single core circuit can be input to the core circuit stage as a common factor.

For example, with reference to FIG. 12, the leak event occurrence rates of a plurality of blocks can be common. Thus, the factor Threshold Leak_rate, which specifies the leak rate that occurs in the STDP threshold counter, and the factor Neuron Leak_rate, which specifies the leak rate that occurs in the ignition counter and the learning counter, can be entered at the core end as a common factor. Also, since the leak event needs to be implemented at a slower rate than other processes, the core circuit may include a separate timer Leak_timer for determining whether to perform a leak event.

STDP_threshold_init indicating the initial threshold value of the STDP threshold counter (STDP_threshold counter), and spike_threshold indicating the ignition threshold value may also be input as common factors in the core circuit.

Inhibition_active, which indicates whether or not to activate the inhibition mechanism, may be input as a common factor. When inhibition_active has a value of 1, the inhibition mechanism is activated. If at least one neuron block of a plurality of neuron blocks in the same core outputs a spike signal, the inhibition mechanism signal inh_event may be input to each of the plurality of neurons as 1, so that all of the plurality of neuron's ignition counters can be reset .

In general inhibition factor General_inhibiton, when a spike signal is output from at least one neuron among a plurality of neurons included in other cores in the same multicore in the multicore block shown in FIG. 15, all the neurons included in the current core Is used to make the inhibition mechanism perform.

On the other hand, signals to be distinguished from each other among a plurality of neurons are a spike signal (spike_out) and a learning request signal (STDP_req) output from each of a plurality of neurons.

The spike signal output conversion unit (Spike 2AER Unit) may output an output AER_out (8 bits) identifying a neuron block outputting the spike signal, based on the spike signal (spike_out) received from each of the plurality of neurons. In addition, the spike signal output conversion unit (Spike 2AER Unit) can output the signal AER_out_v indicating that at least one neuron block has outputted the spike signal.

Also, the STDP arbiter unit (STDP arbiter UNIT) generates a learning request neuron address signal STDP_req_addr, which is a signal for identifying a neuron block outputting the learning request signal, based on the learning request signal STDP_req received from each of the plurality of neurons Can be output. Also, the spike signal output conversion unit (Spike 2AER Unit) may output a signal STDP_event_out indicating that at least one neuron block has output the learning request signal (STDP_req).

When at least one neuron block outputs the learning request signal STDP_req in the same core block, 1 is input as the STDP_envent signal, and the learning counter of all neuron blocks in the same core block is reset. On the other hand, in the multi-core block shown in FIG. 15, when a learning request signal is outputted from at least one neuron among a plurality of neurons included in other cores in the same multicore, the learning counter of all neuron blocks included in the current core is reset .

In FIG. 15, the multicore block includes a plurality of core blocks, and receives common parameters (Parameters_core0, Parameter_core1,? Parameter_core255) shared in each core block.

The merger Merger receives the 8-bit output signal AER_out 8b output from each of the plurality of core blocks, and outputs a 16-bit signal AER_out 16b.

In addition, when at least one neuron block among all the neuron blocks included in at least one core block or a plurality of core blocks among the plurality of core blocks outputs a spike signal, a general suppression factor General_inhibiton for requesting a suppression mechanism is set to 1 Is output.

Similarly, when at least one neuron block among all the neuron blocks included in at least one of the plurality of core blocks or the plurality of core blocks outputs the learning request signal STDP_event_out, the signal STDP_event_in for resetting the learning counter is 1 Value, and is input to all multicore cores.

Although the present disclosure does not define specific hardware and software for implementing blocks within the figures, those skilled in the art will appreciate that the present invention is not limited to particular hardware or software implementations, and that any conventional hardware and / As will be appreciated by those skilled in the art. Accordingly, the hardware may be implemented within one or more application specific integrated circuits, digital signal processors, digital signal processing devices, programmable logic devices, devices, field programmable gate arrays, processors, controllers, microcontrollers, microprocessors, electronic devices and other electronic modules, computers, and combinations thereof. Lt; / RTI &gt;

It is to be understood that the functionality of the components illustrated in the singular components in the description of the present disclosure or in the claims may be implemented by various components of the device when actually implemented, It can be implemented by a single component.

Herein, the terms " first ", " second ", etc. may be used to describe various elements, parts, regions, layers, and / or sections. These terms are used only to distinguish one element from another, unless the context is explicitly stated or the context clearly indicates otherwise, the elements are not limited to these terms. Thus, a first component, component, region, layer, and / or section may be defined as a second component, component, region, layer, and / or section without departing from the scope of the disclosure herein. As used herein, the term " and / or " includes any and all combinations of one or more of the listed items. The elements shown in the singular are not intended to exclude a plurality of elements unless otherwise specified.

In the present application, a method includes one or more steps or operations for implementing the disclosed method. Steps and / or operations may be performed differently than the order described in this disclosure, unless a specific order of steps and / or actions is specified. For example, depending on the associated function, two successively described steps and / or operations may be performed at substantially the same time, or in reverse order.

Claims (20)

A neuron circuit for performing synapse learning on a plurality of synapse weight values,
A synapse weight memory for storing the plurality of synapse weight values;
A first sub-circuit receiving an input signal from a synapse pre-neuron circuit and judging whether or not the received input signal is an activation signal whose synapse value of the received input signal is an active value;
A second sub-circuit for performing a comparison operation between the first cumulative number of times of reception of the activation input signal and the learning threshold value according to the determination result; And
Performing reinforcement learning to set a synapse weight value of at least one input signal received as an active value according to a first probability when the first accumulated reception number reaches the learning threshold value,
And performs weak learning to set each of the plurality of synapse weight values to an inactive value according to a second probability. The method according to claim 1,
Performs a comparison operation between the second cumulative reception count of the activation input signal and the ignition threshold value according to the determination result,
Further comprising a fourth sub-circuit for transmitting a spike signal to a post-synaptic neuron circuit of the neuron circuit when the second cumulative number of received times reaches the ignition threshold value. The method according to claim 1,
Wherein the third sub-
Wherein the reinforcement learning is performed to determine the second probability based on the number of activation synapses in which the synaptic weight in the synaptic weight memory has an active value. The method according to claim 1,
The second sub-
Wherein the learning counter counts the first cumulative number of times of reception and is reset when the first cumulative number of times of reception reaches the learning threshold value. 5. The method of claim 4,
The second sub-
Further comprising a learning threshold counter that counts the learning threshold and increases when the first cumulative number of receptions reaches the learning threshold. 3. The method of claim 2,
Wherein the fourth sub-
And counts the second cumulative number of times of reception and is reset when the second cumulative number of times of reception reaches the ignition threshold value. The method of claim 3,
Wherein the third sub-
A random constant generator,
And performs the reinforcement learning when the positive random number obtained from the random number generator is equal to or less than a reinforcement learning standard constant calculated by multiplying the upper limit value of the random number by the first probability. 8. The method of claim 7,
Wherein the third sub-
For each of the plurality of synapse weight values in the synapse weight memory, if the positive random number constant obtained from the random constant generator is greater than or equal to the weakened learning reference constant calculated by multiplying the upper limit value of the random constant by the second probability, And performs the weakening learning. The method according to claim 1,
Wherein the third sub-
An input event buffer for storing synapse information included in received input signals,
And performs reinforcement learning on a synapse weight value corresponding to each of a predetermined number of synaptic information stored in the input event buffer, in accordance with a reverse order of the stored sequence. A learning system for a plurality of synapse weight values,
Neuron circuit; And
Learning circuit,
Wherein the neuron circuit comprises:
Receiving an input signal from a synaptic neuron circuit,
Determining whether the received input signal is an activation signal whose active value is a synapse value of the received input signal, and transmitting a learning request signal to the learning circuit when the first cumulative reception count of the activation input signal reaches a learning threshold value; In addition,
The learning circuit comprising:
Perform reinforcement learning to set a synaptic weight value of at least one input signal previously received by the neuron circuit to an active value at a first probability upon receiving the learning request signal,
And performs weak learning to set each of the plurality of synapse weight values to an inactive value according to a second probability. A learning method for a plurality of synapse weight values between the neuron circuit and synapse neuron circuits, which is performed by a neuron circuit,
Determining that the received input signal is an active input signal, the synapse value of the received input signal being an active value, upon receiving an input signal from a synapse pre-neuron circuit;
Performing a comparison operation between a first cumulative reception count of the activation input signal and a learning threshold value according to the determination result;
Performing a reinforcement learning step of setting a synaptic weight value of at least one input signal received as an active value according to a first probability when the first cumulative reception count reaches a learning threshold; And
And performing weak learning to set each of the plurality of synapse weight values to an inactive value according to a second probability. 12. The method of claim 11,
Performing a comparison operation between a second cumulative number of times of reception of the activation input signal and the ignition threshold value according to the determination result; And
Further comprising transmitting a spike signal to a post-synaptic neuron circuit of the neuron circuit when the second cumulative number of received times reaches the ignition threshold value. 12. The method of claim 11,
Wherein the weakening learning comprises:
Determining the number of activation synapses having a synaptic weight of the plurality of synapses as a result of the reinforcement learning being performed and determining the second probability based on the determined number of activated synapses . 12. The method of claim 11,
Further comprising the step of initializing the first cumulative number of times of reception if the first cumulative number of receptions reaches a learning threshold. 15. The method of claim 14,
Further comprising increasing the learning threshold if the first cumulative number of receptions reaches the learning threshold. 13. The method of claim 12,
Further comprising the step of initializing the second cumulative number of times of reception if the second cumulative number of receptions reaches the ignition threshold value. 12. The method of claim 11,
Wherein the reinforcement learning step comprises:
And performing the reinforcement learning when the obtained random constant is equal to or less than a reinforcement learning reference constant calculated by multiplying the upper limit value of the random constant by the first probability. 12. The method of claim 11,
Wherein the weakening learning step comprises:
Acquiring a positive random number constant for each of the plurality of synapse weight values and if the obtained random number is equal to or greater than the weak learning standard constant calculated by multiplying the upper limit value of the random constant by the second probability, &Lt; / RTI &gt; 12. The method of claim 11,
Further comprising storing synaptic information included in received input signals,
Wherein the reinforcement learning step comprises:
Further comprising performing reinforcement learning on a synaptic weight value corresponding to each of a predetermined number of previously stored synaptic information, in accordance with a reverse order of the stored sequence. A computer-readable recording medium storing a program for causing a computer to execute the method according to any one of claims 11 to 19.

Images