Rethinking Residual Connection in Training Large-Scale Spiking Neural Networks

Spiking Neural Network (SNN) is known as the most famous brain-inspired model, but the non-differentiable spiking mechanism makes it hard to train large-scale SNNs. To facilitate the training of large-scale SNNs, many training methods are borrowed from Artificial Neural Networks (ANNs), among which deep residual learning is the most commonly used. But the unique features of SNNs make prior intuition built upon ANNs not available for SNNs. Although there are a few studies that have made some pioneer attempts on the topology of Spiking ResNet, the advantages of different connections remain unclear. To tackle this issue, we analyze the merits and limitations of various residual connections and empirically demonstrate our ideas with extensive experiments. Then, based on our observations, we abstract the best-performing connections into densely additive (DA) connection, extend such a concept to other topologies, and propose four architectures for training large-scale SNNs, termed DANet, which brings up to 13.24% accuracy gain on ImageNet. Besides, in order to present a detailed methodology for designing the topology of large-scale SNNs, we further conduct in-depth discussions on their applicable scenarios in terms of their performance on various scales of datasets and demonstrate their advantages over prior architectures. At a low training expense, our best-performing ResNet-50/101/152 obtain 73.71%/76.13%/77.22% top-1 accuracy on ImageNet with 4 time steps. We believe that this work shall give more insights for future works to design the topology of their networks and promote the development of large-scale SNNs. The code will be publicly available.

Spikeformer: A Novel Architecture for Training High-Performance Low-Latency Spiking Neural Network

Spiking neural networks (SNNs) have made great progress on both performance and efficiency over the last few years,but their unique working pattern makes it hard to train a high-performance low-latency SNN.Thus the development of SNNs still lags behind traditional artificial neural networks (ANNs).To compensate this gap,many extraordinary works have been proposed.Nevertheless,these works are mainly based on the same kind of network structure (i.e.CNN) and their performance is worse than their ANN counterparts,which limits the applications of SNNs.To this end,we propose a novel Transformer-based SNN,termed "Spikeformer",which outperforms its ANN counterpart on both static dataset and neuromorphic dataset and may be an alternative architecture to CNN for training high-performance SNNs.First,to deal with the problem of "data hungry" and the unstable training period exhibited in the vanilla model,we design the Convolutional Tokenizer (CT) module,which improves the accuracy of the original model on DVS-Gesture by more than 16%.Besides,in order to better incorporate the attention mechanism inside Transformer and the spatio-temporal information inherent to SNN,we adopt spatio-temporal attention (STA) instead of spatial-wise or temporal-wise attention.With our proposed method,we achieve competitive or state-of-the-art (SOTA) SNN performance on DVS-CIFAR10,DVS-Gesture,and ImageNet datasets with the least simulation time steps (i.e.low latency).Remarkably,our Spikeformer outperforms other SNNs on ImageNet by a large margin (i.e.more than 5%) and even outperforms its ANN counterpart by 3.1% and 2.2% on DVS-Gesture and ImageNet respectively,indicating that Spikeformer is a promising architecture for training large-scale SNNs and may be more suitable for SNNs compared to CNN.We believe that this work shall keep the development of SNNs in step with ANNs as much as possible.Code will be available.