Beyond attention [& transformers]

Hyena is a convolutional layer for LLMs that can shrink the gap with attention, while scaling *subquadratically* in seq len (eg train a lot faster @ 64k + train 100k+ tokens!) 2/

blogs: https://t.co/DIeS1kfyte, https://t.co/FE8BgZYzTX
code: https://t.co/ss9n5bxtDP pic.twitter.com/4yCzbJWlLJ

— Michael Poli (@MichaelPoli6) March 7, 2023

Abstract of the article linked above:

Recent advances in deep learning have relied heavily on the use of large Transformers due to their ability to learn at scale. However, the core building block of Transformers, the attention operator, exhibits quadratic cost in sequence length, limiting the amount of context accessible. Existing subquadratic methods based on low-rank and sparse approximations need to be combined with dense attention layers to match Transformers, indicating a gap in capability. In this work, we propose Hyena, a subquadratic drop-in replacement for attention constructed by interleaving implicitly parametrized long convolutions and data-controlled gating. In recall and reasoning tasks on sequences of thousands to hundreds of thousands of tokens, Hyena improves accuracy by more than 50 points over operators relying on state- spaces and other implicit and explicit methods, matching attention-based models. We set a new state-of- the-art for dense-attention-free architectures on language modeling in standard datasets (WikiText103 and The Pile), reaching Transformer quality with a 20% reduction in training compute required at sequence length 2K. Hyena operators are twice as fast as highly optimized attention at sequence length 8K, and 100× faster at sequence length 64K.