Research On Spike-Driven Learning Algorithm For Neural Network

In recent years,artificial neural networks have become a hot topic in the field of artificial intelligence.Recent studies have shown that spiking neural networks(SNNs)can simulate complex information processing in the brain.There is biological evidence to prove the use of the precise timing of spikes for information coding.However,precise time learning mechanisms in the brain remains an open question.Instruction or supervised learning is a basic way for the brain to acquire new knowledge and develop new skills.Although the concept of supervised learning has been studied for many years,the exact neuron mechanism to achieve this process remain unrevealed.Currently,most existing instructional or learning methods are based on weight adjustments,and the learning accuracy and efficiency of the existing methods still could be improved.In addition,none of them can maintain high accuracy and efficiency under noisy environment.However,noise is widespread in spiking neural networks(SNN).Therefore,it is very meaningful to propose an algorithm that maintain good learning ability even under noisy environment.This thesis is based on these points and aims to propose a more comprehensive algorithm.In general,the content of this thesis are as follows:1.We improve the traditional multi-layer spiking neural network learning algorithm named multi-layer ReSuMe.This thesis first introduces the traditional multi-layer ReSuMe learning algorithm,which is proposed for learning precisely timed spikes in multi-layer spiking neural networks.The multi-layer ReSuMe learning rule has good learning ability,but it still suffers from low efficiency and accuracy.Therefore,to overcome this drawback,an improved Multi-ReSuMe algorithm is proposed to improve the learning accuracy and efficiency.2.A new supervised learning algorithm called Multi-DL-ReSuMe is proposed.The proposed Multi-DL-ReSuMe combines the weight update rule in multi-layer ReSuMe and the proposed delay learning method.Multi-DL-ReSuMe uses more biologically reliable features,so it is more reliable for the simulation of intracerebral information processing.Compared with the original algorithm,combining a delay learning strategy can result in less weight updates,so as to improve the efficiency and accuracy of the algorithm.3.A new supervised learning algorithm,called Multi-DLN-ReSuMe is proposed.The proposed method applies the noise threshold strategy to Multi-DL-ReSuMe.There are two most common errors of spiking neural network under noise environment:trigger an extra spike at undesired time and miss a spike at desired time.The noise threshold uses a dynamic threshold.According to the target spike train,the noise threshold is set before learning.That is,the threshold is set to be slightly higher than the original fixed threshold near the target firing time,and the threshold is set to be lower than the original fixed threshold at the undesired firing time.Therefore,the neurons trained with this dynamic threshold can guarantee that the above two errors can be avoided under noise environment,thereby improving the robustness of the learning algorithm.The experimental results demonstrate that the improved Multi-ReSuMe algorithm and the newly proposed Multi-DL-ReSuMe and Multi-DLN-ReSuMe show better learning efficiency and accuracy than the original Multi-ReSuMe algorithm under different conditions.Furthermore,Multi-DLN-ReSuMe algorithm can still guarantee higher accuracy and efficiency under different intensity noise interference.