Sun Hongze ; Cai Wuque ; Yang Baoxin ; Cui Yan ; Xia Yang ; Yao Dezhong ; Guo Daqing

Abstract

Spiking neural networks (SNNs) have demonstrated excellent capabilities invarious intelligent scenarios. Most existing methods for training SNNs arebased on the concept of synaptic plasticity; however, learning in the realisticbrain also utilizes intrinsic non-synaptic mechanisms of neurons. The spikethreshold of biological neurons is a critical intrinsic neuronal feature thatexhibits rich dynamics on a millisecond timescale and has been proposed as anunderlying mechanism that facilitates neural information processing. In thisstudy, we develop a novel synergistic learning approach that involvessimultaneously training synaptic weights and spike thresholds in SNNs. SNNstrained with synapse-threshold synergistic learning~(STL-SNNs) achievesignificantly superior performance on various static and neuromorphic datasetsthan SNNs trained with two degenerated single-learning models. During training,the synergistic learning approach optimizes neural thresholds, providing thenetwork with stable signal transmission via appropriate firing rates. Furtheranalysis indicates that STL-SNNs are robust to noisy data and exhibit lowenergy consumption for deep network structures. Additionally, the performanceof STL-SNN can be further improved by introducing a generalized joint decisionframework. Overall, our findings indicate that biologically plausible synergiesbetween synaptic and intrinsic non-synaptic mechanisms may provide a promisingapproach for developing highly efficient SNN learning methods.