Optimal ANN-SNN Conversion with Group Neurons

TL;DR

The paper tackles the challenge of high conversion error and long latency in ANN-SNN conversion by introducing Group Neurons (GNs), a neuron type comprised of $\tau$ IF members with intra-GN thresholds and lateral inhibition to increase expressive capacity. They replace IF neurons in the converted SNN with GNs, forming an optimized ANN-SNN conversion framework that preserves ANN accuracy even at very short time-steps. Empirical results on CIFAR-10, CIFAR-100, and ImageNet show GN-based SNNs achieve near-ANN accuracy with drastically reduced latency, outperforming several state-of-the-art conversion methods at small $T$. This work paves the way for practical, high-performance SNN deployments on standard benchmarks and supports efficient neuromorphic computing, with code available at the project repository.

Abstract

Spiking Neural Networks (SNNs) have emerged as a promising third generation of neural networks, offering unique characteristics such as binary outputs, high sparsity, and biological plausibility. However, the lack of effective learning algorithms remains a challenge for SNNs. For instance, while converting artificial neural networks (ANNs) to SNNs circumvents the need for direct training of SNNs, it encounters issues related to conversion errors and high inference time delays. In order to reduce or even eliminate conversion errors while decreasing inference time-steps, we have introduced a novel type of neuron called Group Neurons (GNs). One GN is composed of multiple Integrate-and-Fire (IF) neurons as members, and its neural dynamics are meticulously designed. Based on GNs, we have optimized the traditional ANN-SNN conversion framework. Specifically, we replace the IF neurons in the SNNs obtained by the traditional conversion framework with GNs. The resulting SNNs, which utilize GNs, are capable of achieving accuracy levels comparable to ANNs even within extremely short inference time-steps. The experiments on CIFAR10, CIFAR100, and ImageNet datasets demonstrate the superiority of the proposed methods in terms of both inference accuracy and latency. Code is available at https://github.com/Lyu6PosHao/ANN2SNN_GN.

Figures (3)

• Figure 1: (a) A group neuron composed of three member neurons. (b) Comparison between the traditional ANN-SNN conversion framework and our optimized ANN-SNN conversion framework. In the traditional ANN-SNN conversion framework, the activation functions of the source ANN are typically replaced with IF neurons. Our optimized ANN-SNN conversion framework takes this a step further by replacing IF neurons in the SNN with group neurons.
• Figure 2: Comparison between the firing rates of the IF neuron and the GN ($\tau$=4), at the same time-step setting (T=4).
• Figure 3: Effect of different $\tau$. ResNet-20 on CIFAR100.