Summary of Attention-free Spikformer: Mixing Spike Sequences with Simple Linear Transforms, by Qingyu Wang et al.
Attention-free Spikformer: Mixing Spike Sequences with Simple Linear Transforms
by Qingyu Wang, Duzhen Zhang, Tielin Zhang, Bo Xu
First submitted to arxiv on: 2 Aug 2023
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Spiking Neural Networks (SNNs) are designed using the Transformer architecture, introducing a Spiking Self-Attention (SSA) module that efficiently mixes sparse visual features. The resulting State-Of-The-Art (SOTA) performance is demonstrated on various datasets, outperforming previous SNN-like frameworks. Building upon this foundation, the paper explores replacing the SSA with Linear Transform (LT) modules like Fourier and Wavelet transforms to accelerate Spikformer architecture. This reduces time complexity from quadratic to log-linear, leveraging domain switching between frequency and time domains to extract sparse visual features. The experiment shows that compared to SOTA Spikformer with SSA, Spikformer with LT achieves higher accuracy on neuromorphic datasets (CIFAR10-DVS and DVS128 Gesture) while maintaining comparable performance on static datasets (CIFAR-10 and CIFAR-100). Additionally, Spikformer with LT exhibits 29-51% improvement in training speed, 61-70% in inference speed, and 4-26% reduction in memory usage due to avoiding learnable parameters. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Spiking Neural Networks are a new way to design artificial intelligence. This paper combines SNNs with the Transformer architecture, which is often used for language processing. The result is a super-efficient system that can quickly process images and do tasks like classification. The authors also tested replacing some parts of this system with simpler math operations, which allowed them to speed up the process even more. They did experiments on different types of data and showed that their new approach was better than others in certain situations. |
Keywords
* Artificial intelligence * Classification * Inference * Self attention * Transformer
