Structure Adaptive Learning Method For Spiking Neural Networks

There are numerous interconnected neurons in the human brain,which performs well in learning,memory,decision-making,and so on.Traditional neural networks use ratebased encode to express and calculate information while spiking neural networks use spike-based temporally encode.Due to the complex multi-layer structure and implicit nonlinear mechanism,the formulation of efficient learning methods for spiking neural networks with dynamically adaptive structure is difficult.Therefore,the spiking neural network structure adaptive learning algorithm has important research significance.The main work is summarized as following:(1)We propose a structure adaptive learning method based on pruning strategy for spiking neural networks.We analyze the training process and pruning algorithm of biological neural networks.The error function is calculated by inner products of spike trains.The synaptic weights are modified according to the supervised learning method.In the training process,the hidden neurons are pruned according to the similarity of spike trains.The proposed algorithm is successfully applied to learn spikes sequence learning task.(2)We propose a structure adaptive learning method for multi-layer spiking neural networks based on growth strategies.The main contribution of this work lies in the fact that the proposed algorithm is able to dynamically adjust the connection weights and neurons of the hidden layer.According to the change of learning accuracy,the hidden layer neurons can be recovered.In the spike sequence learning tasks,we analyze the influence of three different adding strategies on the performance of the algorithm.The experimental results verify the effectiveness of the adaptive structure learning method in multi-layer spiking neural network.The adaptive structure networks achieve better performance in the spike train learning task than the fixed structure networks.(3)Medical data sets are of great importance for accurately identifying diseases.We apply the adaptive structure learning algorithm to medical pattern classification problems,such as Wisconsin Breast Cancer data set,Pima Indians Diabetes data set and Diabetic Retinopathy Debrecen data set.The experimental results verify the network of adaptive structure learning algorithm to achieve better performance than fixed structure network.And the results show that the algorithm has good practical applicability.