A pulse neural network optimization method fusing course learning strategies

Abstract

The application discloses a kind of fusion course learning strategy's pulse neural network optimization method, including based on the training sample in training set to the initialized pulse neural network is trained, obtains prediction result;The prediction result is compared with the sample label in training set, and confidence loss is calculated;With confidence loss minimum as objective function, layer by layer forward derivation is carried out, and the parameter of each neuron in pulse neural network is updated;The parameter of optimization pulse neural network is constantly updated, until the iteration of the set round is completed.The application introduces course learning in pulse neural network, dynamically evaluates the difficulty of sample in the training process, expands the contribution of relatively simple sample in back propagation to the current state of model, and reduces the influence on parameter update for the current relatively difficult sample.The pulse neural network optimization strategy has high biological rationality, effectively realizes the process of simulating human learning new knowledge in pulse neural network.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of fusion course learning strategy's pulse neural network optimization method, belong to machine learning, deep learning field. BACKGROUND

[0002] Pulse neural network is the neural network that simulates the way of human brain to process information.Compared with traditional artificial neural network, pulse neural network has the advantages of low energy consumption, efficient processing time series information etc.After development, pulse neural network currently has three main learning ways: supervised learning, unsupervised learning, artificial neural network to pulse neural network.

[0003] In pulse neural network, 0 / 1 discrete pulse signal is transmitted by pulse neuron, for each pulse neuron, when receiving input, neuron membrane potential changes according to neuron dynamics equation.If after receiving input, membrane potential exceeds the specified threshold, neuron outputs 1, fires a pulse, and restores to resting potential;Otherwise, neuron outputs 0, remains silent.This is the charging, discharging and resetting process of pulse neuron.

[0004] And the current pulse neural network model always treats all samples equally in the training process, without considering whether the learning ability of the model under the current state can effectively learn relatively complex and difficult knowledge, which is unreasonable and does not conform to the natural law of human learning, so it has lower biological explainability. SUMMARY

[0005] In order to overcome the defects existing in the prior art, the present application aims to provide a kind of fusion course learning strategy's pulse neural network optimization method.

[0006] The technical scheme provided by the present application to solve the above technical problems is: a kind of fusion course learning strategy's pulse neural network optimization method, comprising the following steps:

[0007] Step S10, randomly initialize each parameter of pulse neural network;

[0008] Step S20, based on training sample in training set, the initialized pulse neural network is trained, and prediction result is obtained;

[0009] Step S30, compare prediction result with sample label in training set, and calculate to obtain training sample cross entropy loss l i ;

[0010] Step S40, cross entropy loss l i Obtained in step S30 is used to calculate the confidence of each training sample σ i ;

[0011] Step S50: Using the cross entropy loss l obtained in step S30 i and the confidence σ obtained in step S40 i , calculate the confidence loss L i ;

[0012] Step S60: The confidence loss L obtained in step S50 is i The minimum is the objective function, and the forward derivative is taken layer by layer to update the parameters of each neuron in each layer of the spiking neural network;

[0013] Step S70, repeat steps S20-S60, continuously update and optimize the parameters of the pulse neural network until the set rounds of iteration are completed, and obtain the pulse neural network model after the pulse neural network is optimized.

[0014] A further technical solution is that the training samples in step S20 are encoded as pulse sequences.

[0015] A further technical solution is that the specific steps of step S20 are:

[0016] Step S21: Encode the training samples in the training set into a pulse sequence and input it into the spiking neural network;

[0017] Step S22, performing calculations backward layer by layer according to the dynamics formula of LIF neurons;

[0018] Step S23: Obtain the model prediction result according to the frequency of pulses emitted by each neuron in the output layer.

[0019] A further technical solution is that the calculation formula in step S40 is:

[0020]

[0021]

[0022] Where: l i is the cross entropy loss; σ i is the confidence level.

[0023] A further technical solution is that the calculation formula in step S50 is:

[0024] L λ (l i ,σ i )=(l i -τ)σ i +λ(logσ i ) 2

[0025] Where: L i is the confidence loss; li is the cross entropy loss; σ i is the confidence level.

[0026] A storage device stores a plurality of programs, wherein the programs are suitable for being loaded and executed by a processor to implement the above-mentioned pulse neural network optimization method integrating the course learning strategy.

[0027] A processing device includes a processor and a storage device; the processor is suitable for executing various programs; the storage device is suitable for storing multiple programs; the programs are suitable for being loaded and executed by the processor to implement the above-mentioned pulse neural network optimization method that integrates the course learning strategy.

[0028] The present invention has the following beneficial effects: the present invention uses the cross entropy loss l of the sample i To evaluate the difficulty of the sample for the model in the current state, the loss l i The larger it is, the more difficult the sample is, and the confidence σ is calculated. i The lower it is, the lower the impact of the difficult sample on parameter update during back propagation. On the contrary, if l i The smaller it is, the higher the credibility of the current sample is, that is, it is given a larger confidence σ i , amplifying its influence on network parameter updates. This process of scaling confidence based on sample difficulty implements the core principle of curriculum learning, that is, the network model learns from simple to difficult, simulating the process of human learning new knowledge, and achieving high biological plausibility of spiking neural networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the pulse neural network structure that introduces the curriculum learning strategy. Implementation Method

[0030] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0031] The present invention provides a pulse neural network optimization method integrating a course learning strategy, comprising the following steps:

[0032] 1. Model training;

[0033] S0: Randomly initialize the parameters of the spiking neural network;

[0034] S1: The training set sample data is encoded as a pulse sequence and input into the network;

[0035] S2: Calculate and process layer by layer according to the dynamics formula of LIF neurons;

[0036] S3: In the output layer, the classification result predicted by the model is obtained according to the frequency of pulses emitted by each neuron in the output layer;

[0037] S4: Compare the prediction results with the sample labels and calculate the cross entropy loss l of the training samples i ;

[0038] S5: Use the cross entropy loss l obtained in S4 i , evaluate the difficulty of each sample, and calculate the confidence σ of each sample according to the formula i ;

[0039]

[0040] in

[0041] S6: cross entropy loss l obtained using S4 i and the confidence σ obtained in S5 i , the new confidence-based loss L is calculated according to the formula i ;

[0042] L λ (l i ,σ i )=(l i -τ)σ i +λ(logσ i ) 2

[0043] S7: so that the confidence loss L obtained in S6 i The minimum is the objective function, and the forward derivative is taken layer by layer to update the parameters of each neuron in each layer;

[0044] S8: Repeat S1-S7, continuously update the parameters of the optimization network, until the set rounds of iteration are completed, and the spiking neural network model after spiking neural network optimization is obtained.

[0045] 2. Model testing:

[0046] S9: The test set sample data is encoded as a pulse sequence and input into the trained spiking neural network.

[0047] S10: Calculation and processing are performed layer by layer according to the dynamic formula of LIF neurons.

[0048] S11: In the network output layer, the classification result predicted by the model is obtained according to the frequency of pulses emitted by each neuron in the output layer.

[0049] S12: Compare the model prediction with the sample label. If they are consistent, the model prediction is correct. Otherwise, the prediction is wrong. Calculate the model classification accuracy.

[0050] This paper introduces curriculum learning into spiking neural networks, dynamically evaluating the difficulty of samples during training. This increases the contribution of samples that are relatively simple for the current state of the model during backpropagation, while reducing the impact of difficult samples on parameter updates. This spiking neural network optimization strategy is highly biologically plausible and effectively simulates the human learning process within spiking neural networks.

[0051] Image classification experiments were conducted on the MNIST, CIFAR10, Fashion-MNIST, and N-MNIST datasets, and the classification accuracy of our CL-SNN model was compared with the existing BackEISNN model results as follows:

[0052] Dataset BackEISNN CL-SNN MNIST 99.67 99.71 CIFAR10 90.93 92.74 Fashion-MNIST 93.45 94.54 N-MNIST 99.57 99.58

[0053] The above description does not limit the present invention in any form. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any technician familiar with the profession can use the technical content disclosed above to make some changes or modifications to equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A pulse neural network optimization method integrating course learning strategy, characterized by: The following steps are involved: Step S10, randomly initializing the parameters of the spiking neural network; Step S20: training the initialized spiking neural network based on the training samples in the training set to obtain a prediction result; Step S30: Compare the prediction results with the sample labels in the training set and calculate the cross entropy loss of the training samples l i ; Step S40: Using the cross entropy loss l obtained in step S30 i , calculate the confidence σ of each training sample i ; Step S50: Using the cross entropy loss l obtained in step S30 i and the confidence σ obtained in step S40 i , calculate the confidence loss L i ; Step S60: The confidence loss L obtained in step S50 is i The minimum is the objective function, and the forward derivative is taken layer by layer to update the parameters of each neuron in each layer of the spiking neural network; Step S70, repeating steps S20-S60, continuously updating and optimizing the parameters of the spiking neural network until the set rounds of iteration are completed, and obtaining the spiking neural network model after the spiking neural network is optimized; Step S80: The test set sample data is encoded into a pulse sequence and input into the trained spiking neural network; Step S90, performing calculations backward layer by layer according to the dynamics formula of LIF neurons; Step S100: In the network output layer, the classification result predicted by the model is obtained according to the frequency of pulses emitted by each neuron in the output layer; Step S110: Compare the model prediction with the sample label. If they are consistent, the model prediction is correct. Otherwise, the prediction is wrong. Calculate the model classification accuracy. The spiking neural network model is applied to image classification.

2. The pulse neural network optimization method integrating course learning strategy according to claim 1 is characterized in that: In step S20, the training samples are encoded as pulse sequences.

3. The pulse neural network optimization method integrating course learning strategy according to claim 1 is characterized in that: The specific steps of step S20 are: Step S21: Encode the training samples in the training set into a pulse sequence and input it into the spiking neural network; Step S22, performing calculations backward layer by layer according to the dynamics formula of LIF neurons; Step S23: Obtain the model prediction result according to the frequency of pulses emitted by each neuron in the output layer.

4. The pulse neural network optimization method integrating course learning strategy according to claim 1 is characterized in that: The calculation formula in step S40 is: Where: l i is the cross entropy loss; σ i is the confidence level.

5. The pulse neural network optimization method integrating course learning strategy according to claim 1 is characterized in that: The calculation formula in step S50 is: L λ (l i ,s i )=(l i -t)s i +λ(log σ i ) 2 Where: L i is the confidence loss; l i is the cross entropy loss; σ i is the confidence level.

6. A storage device storing a plurality of programs, characterized in that: The program is suitable for being loaded and executed by a processor to implement the pulse neural network optimization method integrating course learning strategy as described in any one of claims 1-5.

7. A processing device comprising a processor and a storage device; the processor is adapted to execute various programs; the storage device is adapted to store a plurality of programs; and the processing device comprises: The program is suitable for being loaded and executed by a processor to implement the pulse neural network optimization method integrating course learning strategy as described in any one of claims 1-5.

Images