KR20210146002A - Method and apparatus for training multi-layer spiking neural network - Google Patents

Abstract

Disclosed is a method for training a multilayer spiking neural network. The method repeats a process of generating the input spikes of the current hidden-layer neurons connected to the previous hidden-layer neurons based on the output spike frequency calculated from the previous hidden-layer neurons, performing spike timing dependent plasticity (STDP) learning on the current hidden-layer neurons receiving the input spikes, and storing the output spike frequency calculated in the current hidden-layer neurons for the input spikes, for each of the hidden layers of a multilayer spiking neural network. The method calculates the sum of the minimum synaptic weights related to a speech for each neuron constituting the hidden layers and adjusts the membrane potential threshold voltage of the corresponding neuron based on the sum of the minimum synaptic weights for each neuron, in the multilayer spiking neural network in which the STDP learning for the hidden layers is completed.

Description

Method and apparatus for learning multi-layer spiking neural network {METHOD AND APPARATUS FOR TRAINING MULTI-LAYER SPIKING NEURAL NETWORK}

The present invention relates to a spiking neural network (SNN), and more particularly, to a method and apparatus for learning a multi-layer spiking neural network.

Due to the limitations of the CMOS (Complementary Metal-Oxide-Semiconductor) digital logic circuit, a computing method different from the von Neumann method is being studied. Among them, there is a spiking neural network (SNN) implemented by mimicking the biological characteristics of the brain.

The brain is composed of neurons connected by synapses. A synapse exists at the point where the axon of a neuron and the dendrite of another neuron connect. Referring to the diagram for explaining the spike generation of the neuron of FIG. 1 , the neuron fires when a membrane potential is greater than a threshold voltage, and the fired spike is transmitted through a synapse. SNNs that mimic the spike mechanism of these neurons can operate with low power compared to other neural networks.

A neuron receives signals from other neurons through dendrites, and the membrane potential can be calculated as the sum of the input signals multiplied by the weight of the synapse. The neuron generates a spike voltage with a magnitude of about 100 mV and a short duration of 1 to 2 ms, and the spike is transmitted to the connected neurons through the synapse through the axon of the neuron.

As such, the SNN, which imitates the signal transduction characteristics of synapses and neurons, processes information using spikes fired from each neuron, their timing, synaptic connection relationship, and synaptic weight. To this end, SNNs are generally trained with the Spike Timing Dependent Plasticity (STDP) algorithm. STDP is a learning method that adjusts synaptic weights between neurons based on relative spike timing. However, since the STDP algorithm is mainly used for learning single-layer SNNs, when learning a multi-layer SNN network, the transmission rate of spikes between layers is low, resulting in poor learning efficiency and accuracy.

The task to be solved is to classify the multilayer spiking neural network by layer using STDP learning, learn the synaptic weights, and then adjust the membrane potential threshold voltage of each neuron based on the sum of the synaptic weights related to the firing of each neuron. To provide a learning method and apparatus for a king neural network.

According to an embodiment, a method for a computing device to train a multilayer spiking neural network is provided. The method includes, for each of the hidden layers of the multilayer spiking neural network, generating input spikes of current hidden layer neurons connected to the previous hidden layer neurons based on the output spike frequency calculated in the previous hidden layer neurons, and the input spike After STDP (Spike Timing Dependent Plasticity) learning of the current hidden layer neurons that have received In the multilayer spiking neural network in which the STDP learning is completed, calculating the minimum synaptic weight sum related to the firing for each neuron constituting the hidden layers, and based on the minimum synaptic weight sum for each neuron, the membrane potential threshold voltage of the corresponding neuron including adjusting.

According to an embodiment, it is possible to shorten the learning time of the multilayer spiking neural network and increase the learning accuracy.

According to an embodiment, it is possible to reduce processing time and power consumption of a neuromorphic chip configured with a multilayer spiking neural network.

1 is a diagram for explaining the spike generation of a neuron.
2 is a structural diagram of a learning apparatus according to an exemplary embodiment.
3 is a diagram conceptually illustrating a relationship between a neuron and a synapse of a spiking neural network according to an exemplary embodiment.
4 is a diagram schematically illustrating learning of a multi-layer spiking neural network according to an embodiment.
5 is a flowchart of a method for learning a multi-layer spiking neural network according to an embodiment.
6 is a graph of input spike generation probability according to an exemplary embodiment.
7 is a flowchart of a method for testing a learned multi-layer spiking neural network according to an embodiment.

Hereinafter, with reference to the accompanying drawings, embodiments of the present invention will be described in detail so that those of ordinary skill in the art to which the present invention pertains can easily implement them. However, the present invention may be implemented in several different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification, when a part is "connected" with another part, this includes not only the case of being "directly connected" but also the case of being "electrically connected" with another element interposed therebetween. . Also, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components, unless otherwise stated.

2 is a structural diagram of a learning apparatus according to an exemplary embodiment.

Referring to FIG. 2 , a device for learning a spiking neural network (simply referred to as a “learning device”) 100 may be a computing device operated by at least one processor 110 . Specifically, the learning device 100 includes a processor 110 , a memory 120 , and a storage device 130 , and may further include an input device 140 and an output device 150 , which include a bus 160 . ) can be communicated.

The processor 110 may be a processor of various types that executes the operations of the present invention by processing instructions included in a program, for example, a central processing unit (CPU), a micro processor unit (MPU), a micro controller (MCU). unit), a graphic processing unit (GPU), or the like.

The memory 120 and the storage device 130 store and/or load a corresponding program or various data so that the instructions described to execute the operation of the present invention are processed by the processor 110 . The memory 120 and the storage device 130 may include various types of volatile or non-volatile storage media. For example, the memory 120 may include a read only memory (ROM) 121 and a random access (RAM). memory) 122 . The memory 120 and the storage device 130 may be located inside or outside the processor 110 , and the memory 120 may be connected to the processor 110 through various known means.

The program may include instructions describing an operation of constructing a multi-layer spiking neural network composed of an input layer, a plurality of hidden layers, and an output layer, and learning it according to the present invention.

3 is a diagram conceptually illustrating a relationship between a neuron and a synapse of a spiking neural network according to an exemplary embodiment.

Referring to FIG. 3A , a spiking neural network (SNN) consists of neurons and synapses.

Neurons j is receive input via a synapse an ignition spike X _i from the previous neuron receives the scaled spike of the spike X _i by the synaptic weights W _ij. Neuron j accumulates the product of the input spike and the synaptic weight to the membrane potential, and when the membrane potential is greater than the _{threshold voltage, V th} , it fires and outputs _{the spike Y j .}

_{The spike Y j} fired from neuron j is transmitted to the next neuron through a synapse. In this way, since the amount of signal that transmits the spike of the previous neuron (PRE-neuron) to the next neuron (POST-neuron) is determined by the synaptic weight, the SNN learns the synaptic weight in the learning phase.

On the other hand, since SNN cannot differentiate the membrane potential and synaptic weight of neurons, it cannot be learned by gradient descent and error backpropagation methods commonly used in conventional deep deep learning. Therefore, SNN is learned by the spike timing dependent plasticity (STDP) method, which is a mechanism for changing the strength of synaptic connections in biological synapses.

Biological synapses are neurons It acts as a short-term and long-term memory device due to the chemical action of the transmitter. Short-term memory is implemented by neurotransmitters that remain after the transmission of the spike. Since long-term memory is implemented by the strength of synaptic connections that change due to changes in the number of receptors or synaptic structures, the process of creating long-term memories of synapses using the STDP method, a mechanism for changing the strength of synapses, is called learning.

에 따른 시냅스 연결 강도 변화(Synaptic change)

를 나타내는 STDP 함수이다.

가 양수이고 시각 차이가 작을수록 시냅스 가중치가 커지는 것을 LTP(long-term potentiation)라고 하고,

가 음수이고 시각 차이가 작을수록 시냅스 가중치가 작아지는 것을 LTD(long-term depression)라고 한다. 이와 같이, SNN은 상대적 스파이크 타이밍을 기반으로 뉴런 사이의 시냅스 가중치를 조정하는 STDP 방법으로 학습될 수 있다.Referring to FIG. 3 (b), the visual difference between the post-synaptic spike and the pre-synaptic spike (spike timing)

Synaptic change in strength according to

STDP function representing

is a positive number and the smaller the visual difference, the greater the synaptic weight is called LTP (long-term potentiation),

is a negative number and the smaller the visual difference, the smaller the synaptic weight is called LTD (long-term depression). In this way, the SNN can be learned using the STDP method to adjust the synaptic weight between neurons based on the relative spike timing.

On the other hand, since the STDP method generally uses the relative spike timing between two neurons, there is a limitation in that neurons cannot properly imitate a neural network connected in multiple layers by synapses. That is, in an actual biological neural network, neurons are connected in a multi-layered structure by synapses, and the cerebral cortex of the human brain is composed of approximately 6 to 10 layers. Through this multi-layer structure, starting from simple recognition in the input layer, progressively complex reasoning becomes possible as it goes to the output layer.

In the following, a method for learning an SNN that mimics the multilayer structure of an actual biological neural network will be described in detail.

4 is a diagram schematically illustrating learning of a multilayer spiking neural network according to an embodiment, FIG. 5 is a flowchart of a method for learning a multilayer spiking neural network according to an embodiment, and FIG. 6 is an input spike according to an embodiment This is the generation probability graph.

4 and 5 , the learning apparatus 100 uses the learning data to construct a multi-layer SNN composed of an _{input layer, a plurality of hidden layers (hidden layer 0} , hidden layer ₁ , …, hidden layer _{n-1 ), and an output layer.} learn The learning apparatus 100 learns a multi-layer SNN using the STDP learning method. Each layer of the multilayer SNN is composed of a plurality of nodes, wherein a node corresponds to a neuron, a node-to-node connection corresponds to a synaptic connection between neurons, and a connection strength corresponds to a synaptic weight.

The learning apparatus 100 classifies the multilayer SNN for each layer, and in each hidden layer composed of a plurality of neuron nodes, performs STDP learning of synaptic weights connecting the neuron nodes. Thereafter, the learning apparatus 100 calculates the sum of synaptic weights related to the firing of each neuron based on the learned synaptic weight, and adjusts the membrane potential threshold voltage of the corresponding neuron by using the sum of the synaptic weights corresponding to each neuron. In the multilayered SNN, membrane potential of the hidden layer is _0, so at the same time accumulate accept spike input from a number of nodes in the hidden layer ₀ may spike can occur well. However, since the hidden layers close to the output layer must fire by receiving a spike input from a small number of nodes, the membrane potential threshold voltage for firing the spike must be adjusted to be low.

The learning apparatus 100 generates spikes of input nodes based on learning data corresponding to each input node of the input layer (S110). The learning apparatus 100 repeatedly generates the input spike during the maximum spike frequency through the input spike generator. The input spike may have a Poisson distribution with an occurrence probability proportional to the real value of the input training data.

The learning apparatus 100 inputs the input spikes generated in the input layer to the hidden layer _{0 , which is the first hidden layer, and learns the synaptic weights of neurons constituting the hidden layer 0} during the maximum spike frequency by the STDP learning method (S120). Through STDP learning, the neuron-neuron connection strength is expressed as a real value called synaptic weight. As described in Fig. 3(b), according to the STDP function used in STDP learning, if the pre-synaptic spike occurs before the post-synaptic spike, the spike that occurs in the pre-synaptic neuron affects the spike in the post-synaptic neuron. Therefore, increase the synaptic weight (synaptic connection strength), and if the pre-synaptic spike occurs later than the post-synaptic spike, it is considered that there is no causal relationship between them, and the synaptic weight is decreased.

The learning apparatus 100 stores the frequency of output spikes counted for each neuron of the output spikes generated in each neuron in the _{hidden layer 0} with respect to input spikes generated from one training data (S130). The output spike frequency for each neuron is input as the learning data of the next hidden layer.

The learning apparatus 100 generates input spikes input to the current hidden layer neurons based on the frequency of output spikes for each neuron of the previous hidden layer (S140). The learning apparatus 100 generates an input spike during the maximum spike frequency level through the input spike generator. Learning apparatus 100, the first hidden layer of hidden layer _0, the real number to generate an input spike train of the Poisson distribution from the learning data, and the subsequent hidden layer (hidden in the _first hidden layer _n-1 -) in an integer frequency of the input from the previous hidden layer Generate the input spike train of the Poisson distribution from The learning apparatus 100 may calculate the input spike generation probability of node i as shown in Equation 1 so that the next hidden layer is learned by giving a large weight to the output having the maximum frequency, and Equation 1 is a graph as shown in FIG. can be expressed The learning apparatus 100 may generate an input spike input to the current hidden layer neuron connected to the previous hidden layer neuron according to the frequency of output spikes of the previous hidden layer neuron. That is, the learning apparatus 100 is a neuron of the current hidden layer that receives the spike of the maximum frequency from the previous hidden layer, and inputs the spike generated with the maximum probability. The learning apparatus 100 does not input a spike to a neuron of a current hidden layer connected to a previous hidden layer having an output spike frequency of 0.

The learning apparatus 100 inputs input spikes to the current hidden layer, and learns the synaptic weights of neurons constituting the current hidden layer during the maximum spike frequency by the STDP learning method (S142).

The learning apparatus 100, with respect to the input spike trains, the current hidden layer Output spikes generated from each neuron in the inner circuit store the frequency of output spikes counted for each neuron (S144).

The learning apparatus 100 determines whether individual STDP learning for all hidden layers is completed, and if there are remaining hidden layers, moves to the STDP learning step for each layer ( S140 ) ( S150 ).

When individual STDP learning for all the hidden layers is completed, the learning apparatus 100 selects the neuron having the maximum frequency among the output spike frequencies of the _{last hidden layer n-1, and learns to output the spike of only the selected neuron (Max} Pooling) (S160).

The learning apparatus 100 calculates the sum of synaptic weights related to the firing of neurons constituting the hidden layers in the SNN in which individual STDP learning for all the hidden layers has been completed ( S170 ). The learning apparatus 100 generates an input spike train with training data, inputs it to the STDP learned SNN for each layer, and repeats the process of inputting the input spike train generated based on the spike frequency output from each layer of the SNN to the next layer. While doing so, it is possible to calculate the sum of synaptic weights related to the firing of each neuron. At this time, the learning apparatus 100 calculates the sum of the synaptic weights of each neuron while using the learned synaptic weight and the membrane potential threshold voltage as it is. The sum of the synaptic weights is the sum of the weights of the synapses to which the input spike is applied at the moment each neuron fires.

The learning apparatus 100 extracts the minimum sum of synaptic weights from the sum of synaptic weights of each neuron calculated for the entire learning data (S172).

_{The learning apparatus 100 adjusts the membrane potential threshold voltage (V th} ) of the corresponding neuron based on the sum of the minimum synaptic weights of each neuron constituting the hidden layers ( S180 ). That is, the learning device 100 by a minimum of synaptic weights sum to, even if spikes are inputted, the membrane potential threshold voltage (V _th) to, the over ignition spike to be passed to the next neuron membrane potential threshold voltage (V _th) in each neuron neurons Can be adjusted.

As such, since the learning apparatus 100 increases the spike transmission rate by adjusting the membrane potential threshold voltage of each neuron after individual STDP learning of each hidden layer, the input layer spike is output from the output layer even if it passes through a plurality of hidden layers. SNNs can be trained.

7 is a flowchart of a method for testing a learned multi-layer spiking neural network according to an embodiment.

Referring to FIG. 7 , the learning apparatus 100 generates input spikes corresponding to test data through the input spike generator in the SNN to which the membrane potential threshold voltage for each neuron is adjusted ( S210 ). The input spike generator generates input spikes during the maximum spike frequency magnitude.

The learning apparatus 100 inputs the input spikes generated in the input layer to the first hidden layer, and according to the synaptic weight of the sequentially connected hidden layers and the adjusted membrane potential threshold voltage, calculates the frequency of output spikes for each neuron output from the last hidden layer ( S220).

The learning apparatus 100 inputs the output spike frequency for each neuron of the last hidden layer to the output layer, and obtains the spike of the neuron having the maximum frequency output from the output layer (S230).

As such, according to an embodiment, it is possible to shorten the learning time of the multilayer spiking neural network and increase the learning accuracy. According to an embodiment, it is possible to reduce processing time and power consumption of a neuromorphic chip configured with a multilayer spiking neural network.

Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improved forms of the present invention are also provided by those skilled in the art using the basic concept of the present invention as defined in the following claims. is within the scope of the right.

Claims (1)

A method for a computing device operated by at least one processor to train a multi-layer spiking neural network, comprising:
For each hidden layer of the multilayer spiking neural network, input spikes of current hidden layer neurons connected to the previous hidden layer neurons are generated based on the output spike frequency calculated from the previous hidden layer neurons, and the input spikes are received. Repeating the process of storing the output spike frequency calculated in the current hidden layer neurons with respect to the input spikes after the current hidden layer neurons are trained by STDP (Spike Timing Dependent Plasticity);
In the multilayer spiking neural network in which the STDP learning for the hidden layers is completed, calculating the sum of the minimum synaptic weights related to the firing for each neuron constituting the hidden layers, and
Adjusting the membrane potential threshold voltage of the corresponding neuron based on the sum of the minimum synaptic weights for each neuron
Including, a multi-layer spiking neural network learning method.

Images