Neuronal-Plasticity and Reward-Propagation Improved Recurrent Spiking Neural Networks.

Shuncheng Jia,Xiang Cheng,Tielin Zhang,Bo Xu,Hongxing Liu

doi:10.3389/fnins.2021.654786

Abstract

Different types of dynamics and plasticity principles found through natural neural networks have been well-applied on Spiking neural networks (SNNs) because of their biologically-plausible efficient and robust computations compared to their counterpart deep neural networks (DNNs). Here, we further propose a special Neuronal-plasticity and Reward-propagation improved Recurrent SNN (NRR-SNN). The historically-related adaptive threshold with two channels is highlighted as important neuronal plasticity for increasing the neuronal dynamics, and then global labels instead of errors are used as a reward for the paralleling gradient propagation. Besides, a recurrent loop with proper sparseness is designed for robust computation. Higher accuracy and stronger robust computation are achieved on two sequential datasets (i.e., TIDigits and TIMIT datasets), which to some extent, shows the power of the proposed NRR-SNN with biologically-plausible improvements.

Highlights

Many different types of deep neural networks (DNNs) have been proposed for efficient machine learning on image classification (Ciregan et al, 2012), recognition (Nguyen et al, 2015), memory association (He et al, 2017), and prediction (Kim et al, 2017)
In DNNs, the BP interleaves with feedforward propagation sequentially, and the error signals have to be backpropagated from the output neurons to hidden neurons layer-by-layer, with a risk of gradient disappearance or gradient explosion, especially for extremely-deep networks
This paper focuses more on the research on neuronal dynamics, learning plasticity, and sparseness architectures of spiking neural network (SNN), looking toward a more efficient biologically-plausible computation

Summary

Introduction

Many different types of deep neural networks (DNNs) have been proposed for efficient machine learning on image classification (Ciregan et al, 2012), recognition (Nguyen et al, 2015), memory association (He et al, 2017), and prediction (Kim et al, 2017). The first problem is the increasing number of synaptic parameters. Different types of structures instead of neurons play important roles in different functions of DNNs, where most artificial neurons use a Sigmoid-like activation function for simple non-linear input-output mapping. The unbalanced complexity between artificial neurons and networks allows DNNs to contain a large number of network parameters that can be tuned. In DNNs, the BP interleaves with feedforward propagation sequentially, and the error signals have to be backpropagated from the output neurons to hidden neurons layer-by-layer, with a risk of gradient disappearance or gradient explosion, especially for extremely-deep networks. The dynamic spiking neurons in SNNs are not continuous in the real number field, which is different from the artificial neurons such as the Sigmoid activation function, Tanh activation function, and Rectified linear unit (ReLU).

Objectives

Methods

Findings

Conclusion