Spikformer V2 hits 81.1% on ImageNet, first SNN to break 80%

Spiking Neural Networks (SNNs) have long promised biologically plausible, energy-efficient AI, but their performance lagged behind traditional deep learning. Spikformer V2 from researchers at Peking University and collaborators (Zhaokun Zhou, Yijie Lu, Kaiwei Che, et al.) closes that gap. The team proposes two key innovations: a Spiking Self-Attention (SSA) mechanism that eliminates softmax and operates entirely on spike-based Query, Key, and Value signals, and a Spiking Convolutional Stem (SCS) that adds convolutional layers to the Transformer backbone. The SSA's spike-driven, multiplication-free computation dramatically reduces energy consumption while capturing sparse visual features.

To scale the model, the authors pioneered self-supervised learning (SSL) for SNNs, pre-training Spikformer V2 with a masking-and-reconstruction objective similar to mainstream Vision Transformers, then fine-tuning for ImageNet classification. Results show an 8-layer Spikformer V2 reaches 80.38% accuracy with 4 time steps, while a 172M-parameter, 16-layer version hits 81.10% with only 1 time step—a first for SNNs on ImageNet. This combines Transformer-level accuracy with SNN-level efficiency, opening the door for low-power, event-driven vision systems in edge and robotics applications.