Research map / Architectures and efficiency / Attention and transformers
Fast Transformers with Clustered Attention
Apoorv Vyas, Angelos Katharopoulos, François Fleuret
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add improved clustering (top-k exact keys) on top of clustered attention (ablation) | clustered attention · WSJ test | phone error rate: 7.50 -> 5.61 | Improved |
| use clustered attention instead of full | full attention · WSJ 9 layers | time per epoch (s): 2514 -> 803 | Improved |
| use improved clustered attention instead of full | full attention · WSJ 9 layers | convergence time (h): 87.99 -> 72.14 | Improved |
| use clustered attention instead of Reformer (lsh-4) | Reformer · WSJ test | phone error rate: 8.59 -> 7.50 | Improved |
| use clustered attention instead of full | full attention · Switchboard test | word error rate: 15.0 -> 18.5 | Hurt |
| use improved clustered attention instead of full | full attention · Switchboard test | word error rate: 15.0 -> 15.5 | No clear effect |
| use improved clustered attention instead of full | full attention · Switchboard | convergence time (h): 228.05 -> 127.44 | Improved |
| use clustered attention instead of full | full attention · Switchboard | time per epoch (h): 3.84 -> 1.91 | Improved |
| increase clusters from 100 to 200 at inference | clustered attention (full-trained model) · WSJ val | phone error rate: 44.88 -> 21.76 | Improved |
| add top-k exact keys to clustered attention when approximating full | clustered-100 · WSJ val, full-trained | phone error rate: 44.88 -> 9.29 | Improved |
| keep only the 32 highest-attention keys per query (oracle-top) instead of improved clustered (ablation) | i-clustered · WSJ val, full-trained | phone error rate: 9.29 -> 17.16 | Hurt |
| use improved clustered attention instead of clustered with 25 clusters | clustered-25 · SQuAD | F1: 0.006 -> 0.876 | Improved |
Closest papers by meaning
- Curse of High Dimensionality Issue in Transformer for Long Context Modeling (ICML 2025 · similarity 0.87)
- Theory, Analysis, and Best Practices for Sigmoid Self-Attention (ICLR 2025 · similarity 0.87)
- Attention-based clustering (NeurIPS 2025 · similarity 0.87)
- Mixture-of-Top-$k$ Attention: Efficient Attention as Scalable Fast Weights (NeurIPS 2026 · similarity 0.86)
- kNN Attention Demystified: A Theoretical Exploration for Scalable Transformers (ICLR 2025 · similarity 0.86)
- SEA: Sparse Linear Attention with Estimated Attention Mask (ICLR 2024 · similarity 0.86)
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.
