Spiking Neuromorphic Online Learning System For Epilepsy Diagnosis

The remarkable achievements of deep neural network are accompanied by huge computing requirements and energy consumption,which is difficult to achieve fast response and calculation in real-time scenarios such as smart terminals,mobile devices,and edge computing.The deep spiking neural network combines the computing mode of deep neural network and the information processing method of the spiking neural network,presents and processes information with precisely timed spike trains,and has unique advantages in processing spatiotemporal data such as electroencephalogram.However,the software simulation of deep spiking neural network on the von Neumann computing framework is serial,which cannot take the advantage of high parallelism and low power consumption of deep spiking neural network.By simulating the structure and function of the biological brain,neuromorphic computing attempts to integrate network models and learning algorithms into specialized hardware architectures to achieve more efficient and flexible computing.It is of great research and practical significance to construct deep spiking neural network circuit with online supervised learning capability based on the idea of neuromorphic computing.Currently,most of the supervised learning algorithms for deep spiking neural network are offline mode and involve complicated calculation,which is not suitable to construct hardware circuit of spiking neuromorphic online learning system.From the perspective of software and hardware co-design,this thesis implements a deep spiking neuromorphic online learning hardware system by designing a hardware-friendly algorithm and constructing hardware circuit for the algorithm.The main research work of this thesis includes:(1)A hardware-friendly online supervised learning algorithm is proposed by combining the spike train error calculation and direct feedback alignment mechanism.First,the kernel method transforms the spike train into a continuous function that various over time,which can be used to construct the kernel-based error function of the network.Compared with simple frequency statistics,the kernel method preserves the spatiotemporal information carried in spike trains.Secondly,the direct feedback alignment mechanism directly transmits the output error to each neuron layer of the network through a set of random and fixed matrixes,which simplifies the calculation of the algorithm and relieves the problems of gradient explosion and gradient disappearance caused by error backpropagation layer by layer.Finally,the proposed algorithm is verified by spike train learning task,while the influence of model parameters on the learning performance is analyzed.(2)Based on the proposed hardware-friendly online supervised learning algorithm,a spiking neuromorphic online learning hardware system is designed and implemented.First,by analyzing calculation procedures of basic modules such as spike response neuron and learning algorithm,the corresponding hardware circuits are designed.Secondly,using Vitis HLS of Xilinx,the hardware circuits of the learning system are implemented on the Virtex-7 field programmable gate array chip XC7VX485 T.Finally,on EDA tool Vivado of Xilinx,the performance of the built hardware system is analyzed in terms of hardware resource occupation,power consumption,and running speed.(3)A classification experiment is conducted on epilepsy electroencephalogram dataset CHB-MIT using the built neuromorphic learning system.First,preprocessing such as sample selection,extraction,classification and equal length segmentation is performed on the dataset.Secondly,the BSA algorithm is used to encode the samples generated by preprocessing into corresponding spike trains.Finally,the epilepsy classification experiment is conducted,the accuracies of which are 0.9393 and 0.9275 for training set and testing set,respectively.The experimental results demonstrate that the proposed algorithm and spiking neuromorphic online learning system are effective.In this thesis,hardware and software co-design method is adopted to construct a spiking neuromorphic online learning system,which is used to perform classification experiments on epilepsy electroencephalogram dataset.The research on one hand expands hardware implementation methods of deep spiking neural network with online learning ability,on the other hand,it explores the application of deep spiking neural network in the scenario of intelligent medical care,contributing to promote the research and application of deep spiking neural network.