Reproduce “Towards High-performance Spiking Transformers from ANN to SNN Conversion” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zihan Huang, Xinyu Shi, Zecheng Hao, Tong Bu, Jianhao Ding, Zhaofei Yu, Tiejun Huang

ECMT converts pretrained vision Transformers into spiking networks by correcting nonlinear-operation mismatch through cumulative expectations, and uses multiple firing thresholds to encode more per time step. On ImageNet1k classification, converted EVA reaches 88.60% top-1 accuracy at four steps, versus 89.62% for its ANN, with estimated energy ratio 0.35. These results concern converted pretrained classifiers; energy is modeled, not measured.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.