Hyena Hierarchy: Towards Larger Convolutional Language Models

TL;DR

Hyena introduces a subquadratic convolution-based operator that replaces attention in language models, enabling larger context processing with reduced computational costs and matching the performance of traditional Transformers.

Contribution

The paper presents Hyena, a novel subquadratic operator that replaces attention, allowing for larger sequence modeling with improved efficiency and accuracy.

Findings

Hyena improves accuracy by over 50 points on recall and reasoning tasks.

It achieves state-of-the-art results on language modeling benchmarks.

Hyena operators are significantly faster than traditional attention at large sequence lengths.

Abstract

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 100x faster at sequence length 64K.