AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Laughing Hyena Distillery: Extracting Compact Recurrences From Convolutions

Stefano Massaroli, Michael Poli, Daniel Y Fu, Hermann Kumbong, Rom Nishijima Parnichkun, David W. Romero, Aman Timalsina, Quinn McIntyre, Beidi Chen, Atri Rudra, Ce Zhang, Christopher Re and 2 more

NeurIPS 2023 · Architectures and efficiency › Attention and transformers · approach: State space and recurrent sequence models · 4 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use multi-head Hyena (8 heads, tied long filters) instead of HyenaHyena 153M · The Pile, 5B tokensperplexity: 13.3 -> 12.1Improved
use multi-head Hyena instead of HyenaHyena 153M · The Pile, 15B tokensperplexity: 11.1 -> 10.6Improved
use multi-head Hyena instead of HyenaHyena 355M · The Pile, 15B tokensperplexity: 9.2 -> 8.9Improved
use multi-head Hyena instead of Hyena, full 300B-token trainingHyena small · The Pile, 300B tokensperplexity: 9.3 -> 8.7Improved
distill MultiHyena filters to state order 16 (ablation)MultiHyena 154M · LAMBADAaccuracy: 43.2 -> 43.1No clear effect
distill MultiHyena filters to state order 16 (ablation)MultiHyena 154M · PIQAaccuracy: 64.6 -> 64.7No clear effect
distill MultiHyena filters to state order 8 instead of 16 (ablation)MultiHyena 154M · LAMBADAaccuracy: 43.1 -> 0.0Hurt
distill MultiHyena filters to state order 8 instead of 16 (ablation)MultiHyena 154M · PIQAaccuracy: 64.7 -> 51.5Hurt
distill to state order 4 instead of 8 (ablation)MultiHyena 154M · HellaSwagnormalized accuracy: 32.7 -> 26.4Hurt
distill to order 8 instead of 16 (ablation)MultiHyena 154M · Winograndeaccuracy: 52.6 -> 51.8Mixed
increase SSM state dimension from 32 to 64 (ablation)LaughingHyena · peak throughputthroughput: 2% reductionHurt
replace Transformer with distilled recurrent Hyena at 1.3BLaughingHyena 1.3B · peak generation throughputthroughput: 10x higherImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy