Deep Supervised Learning Algorithm Of Spiking Neural Network Based On Back Reconstruction

Deep neural networks,with their complex network structures,are powerful computational tools for solving pattern recognition,function computation,and classification problems.They are capable of learning and representing more complex and abstract data features than shallow networks.They are widely used for solving complicated pattern recognition,function computation,and classification problems.In recent years,the development trend of neural computing models has been towards the direction of deep spiking neural network models based on spike train encoding.This type of model is more closely aligned with the working mechanism of the biological nervous system and can effectively handle spatiotemporal problems.The construction of deep spiking neural network learning algorithms is an important challenge in the research of brain-like computing,and its key lies in considering the precise temporal characteristics of spike neural information.The widespread application of traditional deep neural networks is due to the error backpropagation learning algorithm based on gradient descent.The key to implementing error backpropagation is to obtain the network expectation and calculate the network error.The total network error is fed back to the upstream neurons of the network through gradient descent and the weight adjustment is realized layer by layer.(1)In the biologically-inspired synaptic plasticity model,the adjustment of network weights is only influenced by the activity of local presynaptic and postsynaptic spike trains.Inspired by the mechanism of synaptic plasticity and error backpropagation,this thesis proposes a spike back reconstruction algorithm.The back reconstruction algorithm calculates the local expectations and errors of each layer of the network by backpropagating the expected output signal through the network.According to the principle of gradient descent,the local error signal is fed back to the previous layer neurons to achieve the training of the deep spiking neural networks.The effectiveness of the back reconstruction mechanism has been verified through spike train experiments,and the impact of various parameters on the algorithm’s learning accuracy and convergence speed has been analyzed.In addition,this thesis compares and analyzes the performance differences between the back reconstruction mechanism and the backpropagation mechanism in the networks with different scales.(2)The layer-wise expectation calculation method in back reconstruction learning algorithm lacks locality and biological plausibility.To solve this problem,this thesis proposes an optimized spike broadcast reconstruction learning algorithm.The broadcast reconstruction algorithm generates desired spike trains for each layer through a random direct path,thereby avoiding layer-wise transmission of backwards stimulus signals.The correctness and rationality of the broadcast reconstruction learning algorithm are verified by spike train learning experiments.The broadcast reconstruction mechanism has been experimentally validated for its effectiveness through spike train learning experiments,and the impact of different parameters on the algorithm’s learning accuracy and convergence speed has been analyzed.In addition,this thesis compares and analyzes the performance differences of broadcast reconstruction mechanism and back reconstruction mechanism under different network scales.(3)This study examined the data analysis capabilities of the proposed back reconstruction and broadcast reconstruction learning algorithms in real-world tasks,using cognitive workload assessment based on electroencephalography(EEG).The EEG dataset used in this study is the EEG During Mental Arithmetic Tasks(DMAT)dataset.According to the DMAT dataset,two cognitive workload analysis tasks can be constructed,namely,the cognitive workload detection task and the cognitive workload ranking task.The back reconstruction learning algorithm achieved an average classification accuracy of 97.22% and 95.55% on the two cognitive workload analysis tasks,respectively.The broadcast reconstruction learning algorithm achieved classification accuracy of 97.22% and 99.44% on the two tasks,respectively.