LASER: Attention using Exponential Transformation
Keywords: Attention Mechanism, LLM, Transformer, Conformer, ViT
Abstract: Transformers have had tremendous impact for several sequence related tasks. The softmax based dot-product attention mechanism plays a key role in the Transformer's ability to retrieve from any part of the sequence via a parameterized query-key-value mechanism. However, the softmax operation can backpropagate small gradients thus inhibiting learning. In this paper, we fix this by introducing a new attention mechanism called LASER attention, which admits a log-sum-exp structure and propagates a larger gradient signal. We show that LASER attention can be implemented by making small modifications to existing attention implementations. We conduct experiments on large language models (LLMs) with upto 2.2 billion parameters where we show improvements of upto 3.38\% and 1\% on an average compared to standard attention on downstream one-shot evaluations. We also evaluate on transformers spanning different modalities (vision, speech and text): Vision Transformer (ViT) on Imagenet (1.2\% improvement in accuracy), Conformer on the Librispeech speech-to-text task (2.25\% relative improvement) and encoder-only BERT Transformer with 2.2 billion parameters (0.93\% relative improvement).
Primary Area: foundation or frontier models, including LLMs
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12636
Loading