MT-SNN: Enhance Spiking Neural Network with Multiple Thresholds

TL;DR

This paper introduces Multiple Threshold (MT) approaches to significantly enhance SNN accuracy by mitigating precision loss, and proposes two distinct modes for MT implementation, depending on the membrane update rule: parallel mode and cascade mode.

Abstract

Spiking neural networks (SNNs) present a promising energy efficient alternative to traditional Artificial Neural Networks (ANNs) due to their multiplication-free operations enabled by binarized intermediate activations. However, this binarization leads to precision loss, hindering the SNN performance. In this paper, we introduce Multiple Threshold (MT) approaches to significantly enhance SNN accuracy by mitigating precision loss. We propose two distinct modes for MT implementation, depending on the membrane update rule: parallel mode and cascade mode. MT-SNN models can be efficiently trained on standard hardwares like GPUs and TPUs, while retaining the multiplication-free advantage crucial for deployment on neuromorphic devices. Our extensive experiments on CIFAR10, CIFAR100, ImageNet, and DVS-CIFAR10 datasets demonstrate that both MT modes substantially improve the performance of single-threshold SNNs, achieving higher accuracy with fewer time steps and comparable energy consumption. Moreover, MT-SNNs outperform state-of-the-art (SOTA) results. Notably, with MT, a Parametric-Leaky-Integrate-Fire (PLIF) based ResNet-34 architecture reaches 72.17\% accuracy on ImageNet with a single time step, surpassing the previous SOTA by 2.75\% despite using 4 steps.

Figures (6)

• Figure 1: Spikes with single threshold and multiple threshold, $\mu$ is membrane and $S$ is output spike, $th$ is short for threshold. (a): Single Threshold (ST): Spikes are fired only if the membrane exceeds $\mu_{th}$. (b): PARALLEL MT: Three independent thresholds are applied, thus three spike sequences are generated and merged. (c): CASCADE MT: Three thresholds are applied and will be checked sequentially from highest to lowest, the fired spikes are weighted before aggregation.
• Figure 2: Performance comparison of ST and MT on CIFAR10 / CIFAR100 with various steps. There are four arms in each figure, namely ANN, SNN with Single Threshold(ST), SNN with PARALLEL MT, CASCADE MT.
• Figure 3: Performance comparison of ST and PARALLEL MT on CIFAR10-DVS with various steps.
• Figure 4: Ablation study of PARALLEL MT and CASCADE MT on CIFAR100. (a): Impact of thresholds $\Delta$ in PARALLEL MT. (b): Impact of layer placement in PARALLEL MT. (c): Impact of bit number in CASCADE MT.
• Figure 5: Activation Sparsity of VGG-9 layers on CIFAR100