Improving multi-layer spiking neural networks by incorporating brain-inspired rules

Abstract

This paper introduces seven brain-inspired rules that are deeply rooted in the understanding of the brain to improve multi-layer spiking neural networks (SNNs). The dynamics of neurons, synapses, and plasticity models are considered to be major characteristics of information processing in brain neural networks. Hence, incorporating these models and rules to traditional SNNs is expected to improve their efficiency. The proposed SNN model can mainly be divided into three parts: the spike generation layer, the hidden layers, and the output layer. In the spike generation layer, non-temporary signals such as static images are converted into spikes by both local and global feature-converting methods. In the hidden layers, the rules of dynamic neurons, synapses, the proportion of different kinds of neurons, and various spike timing dependent plasticity (STDP) models are incorporated. In the output layer, the function of classification for excitatory neurons and winner take all (WTA) for inhibitory neurons are realized. MNIST dataset is used to validate the classification accuracy of the proposed neural network model. Experimental results show that higher accuracy will be achieved when more brain-inspired rules (with careful selection) are integrated into the learning procedure.

创新点

本文总结和归纳了七条受脑启发的学习准则,并应用于改善脉冲神经网络。这些学习准则都来源于对生物脑的实验研究,并各自从不同的侧面反映了生物网络的学习特性,如神经元的动态分配、突触的自适应生长和消亡机制、不同的突触可塑性学习机制(如不同类型的时序依赖突触可塑性)、网络背景噪声对学习的调控机制、兴奋性和抑制性神经元的比例对学习的调节机制等。本文通过组合上述不同的受脑启发的规则,通过实验研究验证了:随着越来越多的、经过仔细选择的、受脑启发的规则的引入,深层脉冲神经网络能够得到越来越好的分类性能。