Spotlight Poster
QKFormer: Hierarchical Spiking Transformer using Q-K Attention
chenlin zhou · Han Zhang · Zhaokun Zhou · Liutao Yu · Liwei Huang · Xiaopeng Fan · Li Yuan · Zhengyu Ma · Huihui Zhou · Yonghong Tian
East Exhibit Hall A-C #3702
Spiking Transformers, which integrate Spiking Neural Networks (SNNs) with Transformer architectures, have attracted significant attention due to their potential for low energy consumption and high performance. However, there remains a substantial gap in performance between SNNs and Artificial Neural Networks (ANNs). To narrow this gap, we have developed QKFormer, a direct training spiking transformer with the following features: i) Linear complexity and high energy efficiency, the novel spike-form Q-K attention module efficiently models the token or channel attention through binary vectors and enables the construction of larger models. ii) Multi-scale spiking representation, achieved by a hierarchical structure with the different numbers of tokens across blocks. iii) Spiking Patch Embedding with Deformed Shortcut (SPEDS), enhances spiking information transmission and integration, thus improving overall performance. It is shown that QKFormer achieves significantly superior performance over existing state-of-the-art SNN models on various mainstream datasets. Notably, with comparable size to Spikformer (66.34 M, 74.81\%), QKFormer (64.96 M) achieves a groundbreaking top-1 accuracy of 85.65\% on ImageNet-1k, substantially outperforming Spikformer by 10.84\%. To our best knowledge, this is the first time that directly training SNNs have exceeded 85\% accuracy on ImageNet-1K.
Live content is unavailable. Log in and register to view live content