SpikeZIP-TF: Conversion is All You Need for Transformer-based SNN

Brain-inspired spiking neural network (SNN) has attracted great attention because of its high efficiency over the traditional artificial neural network (ANN). Currently, ANN to SNN conversion methods can produce SNNs using convolution neural network as backbone architecture which achieves on-par accuracy w.r.t ANNs with ultra-low latency (e.g., 8 time-steps) on computer vision (CV) tasks. Although Transformerbased networks have achieved the prevailing precision on both CV and natural language processing (NLP) tasks, Transformer-based SNNs are still lagging behind their ANN counterparts. In this work, we introduce a novel ANN-to-SNN conversion method, called SpikeZIP-TF, through which ANN and the converted SNN are exactly equivalent thus incurring no accuracy degradation. SpikeZIP-TF achieves 83.82% Top-1 accuracy on the CV image classification task with ImageNet dataset and 93.79% accuracy on the NLP dataset (SST-2), which both are higher than SOTA Transformer-based SNNs. The code is publicly available at: https://github.com/Intelligent-Computing-Research-Group/SpikeZIP-TF