Supervised Learning Algorithm Of Deep Spiking Neural Networks Based On Feedback Alignment

With the rapid development of the artificial neural networks,biologically in-spired spiking neural networks have received more and more attention in recent years.Spiking neural networks use spike trains to represent and process information based on more bioauthentic spike neuron models.This encoding method integrates multiple aspects of information,such as time,frequency,and phase.The biological brain has a multi-level network structure,which can transform and abstract the characteristics of signals layer by layer.According to this idea,researchers have proposed neural networks with deep structure and their deep learning methods.It is a challenging task in the research of machine learning to construct the deep learning method of spiking neural network which is widely applicable by combining the information processing mode of spiking neural network and the calculation mode of deep neural network.However,due to the non-differentiable nature of the internal state of spiking neurons,the differential-based optimization methods which has been successfully applied to the traditional neural network is no longer suitable for the spiking neural network.Although spiking networks are theoretically considered to have higher computational power than traditional neural network,it remains a challenge to train deep spiking networks.Some spiking neural networks implement deep learning using backpropagation,but it is obvious that they still cannot avoid the biologically unreasonable weight transport problem of backpropagation algorithms.More deep spiking neural networks are converted from traditional neural networks,but during the conversion process,a series of problems such as reduced accuracy,neurons under-activated or over-activated,negative values cannot be represented and increase network complexity will inevitably occur.In view of the above-mentioned shortcomings in deep spiking neural networks,this paper proposes two supervised learning algorithms.Firstly,according to the definition and nature of the inner product of spike trains,a supervised learning algorithm of the deep spiking neural network based on the inner product of spike trains and its simplified algorithm are proposed.Using the inner product of the spike train to construct the error function and synaptic weight learning rules.In view of the computational complexity of the algorithm,we have further simplified its learning rules.We performed a series of spike trains learning experiments on both algorithms.We analyze the learning process of the spike trains to verify the effectiveness of the proposed algorithm.We use different learning rates,different spike firing rates,and different upper limits of the learning epochs to analyze the influence of network parameters on the performance of the algorithm.The experimental results show that the proposed algorithm has high learning accuracy and is very effective in dealing with complex spatio-temporal pattern learning problems.Secondly,according to the back propagation mechanism of spike train errors,a supervised learning algorithm for deep spiking neural networks based on the feedback alignment mechanism and its simplified algorithm are proposed.We use the feedback alignment mechanism to construct the algorithm's learning rules based on the inner product of the spike trains,and construct a simplified learning algorithm based on the broadcast alignment mechanism(a variant of the feedback alignment mechanism).We demonstrate the learning process of spike trains learning to verify the effectiveness of the algorithm,and use different upper limits of the random fixed matrix,different spike firing rates and different length of spike trains to analyze the performance of the algorithm.Experimental results show that the proposed algorithm and its simplified algorithm are correct and effective,and can learn complex spatio-temporal patterns well.Finally,the two proposed algorithms and their simplified algorithms are applied to the nonlinear pattern classification problem.In the experiments,we selected Wisconsin Diagnostic Breast Cancer Data Set and Audit data with more attributes in the UCI standard dataset to evaluate the performance of the proposed algorithm.The experimental results show that the two proposed algorithms and their simplified algorithms can effectively solve the problem of nonlinear pattern classification,and have certain practical value.