LMS: Learnable Maximum Spike with Optimal Spike Representation for High-Performance and Efficient Spiking Neural Networks

Binghao Ye, Wenjuan Li, Bing Li, Dong Liang, Weiming Hu, Kun Shang

Keywords: Spiking Neural Networks, Learnable Maximum Spike, Optimal Spike Representation

TL;DR: Dynamically learning spike maximum of neuron and solving integer programming problems to minimize inference energy consumption

Abstract: Spiking Neural Networks (SNNs) have garnered increasing attention due to their brain-inspired mechanisms. By encoding information with sparse binary spikes, they replace multiplications with additions, substantially reducing energy consumption. However, the binary spike emission inherently leads to significant information loss. In this work, we propose a Learnable Maximum Spike (LMS) neuron, which emits integer values during training and dynamically learns the maximum membrane potential for each layer based on its own membrane potential distribution, thereby determining the maximum integer value the neuron can emit (referred to as spike maximum). Additionally, we introduce a decay balancing coefficient that allows the spike maximum to adapt to the gradient and change in membrane potential distribution between the early and late stages of training, thereby further enhancing the network's performance. Finally, to preserve spike-driven inference, we transform the binary representation problem of emitted values into an integer programming problem, yielding an optimal spike representation of integers that minimizes energy consumption. Extensive experiments have validated the effectiveness of the proposed LMS neuron, which consistently outperforms current state-of-the-art methods on static datasets (CIFAR10, CIFAR100, ImageNet) and a neuromorphic dataset (CIFAR10-DVS). Furthermore, LMS requires less inference memory (**-7.16**\%), shorter inference time (**-12.09**\%), and lower energy consumption (**-61.9**\%).

Primary Area: applications to neuroscience & cognitive science

Submission Number: 3264