SAFformer: Novel Spiking Transformer Achieves 80.50% Top-1 Accuracy on ImageNet-1K

A new architecture called SAFformer has been developed by researchers, representing a Spiking Transformer that utilizes an active predictive filtering approach, drawing inspiration from the brain's predictive coding system. In contrast to traditional Spiking Transformers that operate on a passive reactive basis, SAFformer actively minimizes predictable signals, honing in on important visual elements and thereby decreasing the computational load associated with redundant visual information. This model sets a new benchmark for performance on the CIFAR-10/100 and CIFAR10-DVS datasets. It also achieves 80.50% Top-1 accuracy on ImageNet-1K, utilizing just 26.58M parameters with minimal energy use. This advancement addresses a key limitation in existing Spiking Transformers, allowing them to focus on pertinent task information. The research is published on arXiv, reference 2605.08270.