Design And FPGA Based Implementation Of Liquid State Machine For Epilepsy Seizure Eeg Signal Classification

Liquid state machine(LSM)is a type of compact spiking neural network with recurrent connections,which has gained extensive research and attention due to its high computational power,biological rationality,simple structure and low training complexity,especially for current energy-efficient information processing and computing at the edge of Io T.In this thesis,for automatic classification of epileptic seizures on electroencephalogram(EEG)data,we investigate the optimization of topology and computational parameters of LSM architecture as well as its hardware implementation of FPGA to achieve high classification accuracy.LSMs with cubic structure were constructed based on the Brian 2 spiking neuron network simulator with leaky integrate and fire(LIF)neuron model.The optimization of topology and nonlinear dynamics of the LSM was investigated using particle swarm optimization(PSO)algorithm by searching the hyperparameters of the LSM such as input layer connection probability and membrane time constant of spiking neurons in liquid layer,combined with the Softmax classifier for the seizure three-category classification task.The effect of inertia weights in the PSO algorithm on the classification performance was studied and its improvement was evaluated by principal component analysis and kernel quality of the liquid layer in LSM.The LSM of 4 × 4 × 10 achieved an optimal average accuracy of 95%by 10-fold cross-validation,which has 23.7% improvement in accuracy compared to other similar work.A grid search was performed around the hyperparameters of the highest performance LSM to provide parameter bounds for its FPGA hardware implementation.The Xilinx Virtex-7 VC707 FPGA evaluation board was used to design the liquid layer and its modules such as state extraction and input/output interaction with the host based on the Verilog hardware description language.The best accuracy of 91 % for the epilepsy seizure EEG signal three-category classification task was obtained with the software Softmax classifier.The digital sub-system ran at a clock frequency of 50 MHz with a total system power of 1.091 W.The difference between the accuracy of the FPGA-implemented LSM and the software simulation came from that between the fixed-point and floating-point operations,as shown by the Euclidean distance and cosine similarity of their state spaces.An effort was also done to reduce the dimension of the liquid layer from both software simulation and hardware implementation.