Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention (AI Paper Explained)
Details
#transformer #nystromer #nystromformer The Nyströmformer (or Nystromformer, Nyströmer, Nystromer), is a new drop-in replacement for approximating the Self-Attention matrix in Transformers with linear memory and time requirements. Most importantly, it uses the Nystrom-Method to subselect (or segment mean) queries and keys as so-called landmarks and uses those to reconstruct the inherently low-rank attention matrix. This is relevant for many areas of Machine Learning, especially Natural Language processing, where it enables longer sequences of text to be processed at once. OUTLINE: 0:00 - Intro & Overview 2:30 - The Quadratic Memory Bottleneck in Self-Attention 7:20 - The Softmax Operation in Attention 11:15 - Nyström-Approximation 14:00 - Getting Around the Softmax Problem 18:05 - Intuition for Landmark Method 28:05 - Full Algorithm 30:20 - Theoretical Guarantees 35:55 - Avoiding the Large Attention Matrix 36:55 - Subsampling Keys vs Negative Sampling 43:15 - Experimental Results 47:00 - Conclusion & Comments Paper: https://arxiv.org/abs/2102.03902 Code: https://github.com/mlpen/Nystromformer Appendix: https://github.com/mlpen/Nystromformer/blob/main/doc/Nystromformer_Supplement.pdf LRA Results: https://twitter.com/tanmingxing/status/1359301186734620675 Twitter lucidrains w/ author: https://twitter.com/lucidrains/status/1359597104075661312 Twitter lucidrains w/ _clashluke: https://twitter.com/_clashluke/status/1359483460851802115 Abstract: Transformers have emerged as a powerful tool for a broad range of natural language processing tasks. A key component that drives the impressive performance of Transformers is the self-attention mechanism that encodes the influence or dependence of other tokens on each specific token. While beneficial, the quadratic complexity of self-attention on the input sequence length has limited its application to longer sequences -- a topic being actively studied in the community. To address this limitation, we propose Nyströmformer -- a model that exhibits favorable scalability as a function of sequence length. Our idea is based on adapting the Nyström method to approximate standard self-attention with O(n) complexity. The scalability of Nyströmformer enables application to longer sequences with thousands of tokens. We perform evaluations on multiple downstream tasks on the GLUE benchmark and IMDB reviews with standard sequence length, and find that our Nyströmformer performs comparably, or in a few cases, even slightly better, than standard Transformer. Our code is at this https URL. Authors: Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh Links: TabNine Code Completion (Referral): http://bit.ly/tabnine-yannick YouTube: https://www.youtube.com/c/yannickilcher Twitter: https://twitter.com/ykilcher Discord: https://discord.gg/4H8xxDF BitChute: https://www.bitchute.com/channel/yannic-kilcher Minds: https://www.minds.com/ykilcher Parler: https://parler.com/profile/YannicKilcher LinkedIn: https://www.linkedin.com/in/yannic-kilcher-488534136/ BiliBili: https://space.bilibili.com/1824646584 If you want to support me, the best thing to do is to share out the content :) If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this): SubscribeStar: https://www.subscribestar.com/yannickilcher Patreon: https://www.patreon.com/yannickilcher Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2 Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

0:00 - Intro & Overview 2:30 - The Quadratic Memory Bottleneck in Self-Attention 7:20 - The Softmax Operation in Attention 11:15 - Nyström-Approximation 14:00 - Getting Around the Softmax Problem 18:05 - Intuition for Landmark Method 28:05 - Full Algorithm 30:20 - Theoretical Guarantees 35:55 - Avoiding the Large Attention Matrix 36:55 - Subsampling Keys vs Negative Sampling 43:15 - Experimental Results 47:00 - Conclusion & Comments
Comments
loading...