DC Field | Value | Language |
---|---|---|
dc.contributor.author | Ahn, Kwangjun | ko |
dc.contributor.author | Cheng, Xiang | ko |
dc.contributor.author | Song, Minhak | ko |
dc.contributor.author | Yun, Chulhee | ko |
dc.contributor.author | Jadbabaie, Ali | ko |
dc.contributor.author | Sra, Suvrit | ko |
dc.date.accessioned | 2024-03-13T13:00:15Z | - |
dc.date.available | 2024-03-13T13:00:15Z | - |
dc.date.created | 2024-03-13 | - |
dc.date.issued | 2024-05-07 | - |
dc.identifier.citation | 12th International Conference on Learning Representations, ICLR 2024 | - |
dc.identifier.uri | http://hdl.handle.net/10203/318546 | - |
dc.description.abstract | Transformer training is notoriously difficult, requiring a careful design of optimizers and use of various heuristics. We make progress towards understanding the subtleties of training Transformers by carefully studying a simple yet canonical linearized shallow Transformer model. Specifically, we train linear Transformers to solve regression tasks, inspired by J. von Oswald et al. (ICML 2023), and K. Ahn et al. (NeurIPS 2023). Most importantly, we observe that our proposed linearized models can reproduce several prominent aspects of Transformer training dynamics. Consequently, the results obtained in this paper suggest that a simple linearized Transformer model could actually be a valuable, realistic abstraction for understanding Transformer optimization. | - |
dc.language | English | - |
dc.publisher | International Conference on Learning Representations (ICLR) | - |
dc.title | Linear attention is (maybe) all you need (to understand Transformer optimization) | - |
dc.type | Conference | - |
dc.type.rims | CONF | - |
dc.citation.publicationname | 12th International Conference on Learning Representations, ICLR 2024 | - |
dc.identifier.conferencecountry | AU | - |
dc.identifier.conferencelocation | Vienna | - |
dc.contributor.localauthor | Yun, Chulhee | - |
dc.contributor.nonIdAuthor | Ahn, Kwangjun | - |
dc.contributor.nonIdAuthor | Cheng, Xiang | - |
dc.contributor.nonIdAuthor | Song, Minhak | - |
dc.contributor.nonIdAuthor | Jadbabaie, Ali | - |
dc.contributor.nonIdAuthor | Sra, Suvrit | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.