Few Heads are Enough

Róbert Csordás, Piotr Piękos, Kazuki Irie, Jürgen Schmidhuber

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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.

Keywords: transformers, attention, moe, mixture of experts, efficient transformers, language modelling

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We show that with MoE value and output projections it is possible to use less heads, reducing resource usage of transformers.

Abstract:

The costly self-attention layers in modern Transformers require memory and compute quadratic in sequence length. Existing approximation methods usually underperform and fail to obtain significant speedups in practice. The recently proposed Flash-Attention reduces both compute and memory through a hardware-aware implementation. Can we achieve this also through algorithmic improvements? Here we present Expert Projection Attention (EPA) - a novel method that reduces both compute and memory requirements, while matching the language modeling performance of baseline Transformers using the same parameter budget. EPA uses Mixture-of-Experts (MoE) layers for the value and output projections and requires 4 to 8 times fewer attention matrices than standard Transformers. Our novel attention can also be combined with MoE MLP layers, resulting in an efficient "Fast Transformer".

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: 9177