Research map / Architectures and efficiency / Attention and transformers
A Capacity-Based Rationale for Multi-Head Attention
Micah Adler
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 |
|---|---|---|---|
| split fixed total key dimension across more heads (multi-head instead of single head) (ablation) | max-over-heads attention on permutation graphs · permutation graphs m=512 d_model=64 / m=256 d_model=32 | micro-F1 reaching 0.99 | Improved |
| use multiple heads in a full Transformer block with frozen GPT-2 embeddings (ablation) | single-layer Transformer, m=6144 · permutation-matching retrieval | test accuracy: h=1 saturates near 0.81, with some seeds collapsing to 0 | Improved |
| use many heads (h=8) at very small D_K=64 (d_k=8) (ablation) | single-layer Transformer, m=6144 · permutation-matching retrieval | test accuracy | Hurt |
| add multiple heads under message retrieval with value channel (d_model=16, h>=4 vs h=1) (ablation) | value-channel attention, m=64 · message retrieval | test MSE | Improved |
| relax compression to d_model=32 (multi-head benefit shrinks) (ablation) | value-channel attention, m=64 · message retrieval | test MSE | No clear effect |
| add softmax with sqrt(d_k) scaling and summed heads (ablation) | idealized attention model · permutation graphs | D_K needed for F1 0.99 | Hurt |
| increase graph density from permutation (r=1) to r-regular graphs (ablation) | max-over-heads attention · r-regular graphs m=128 d_model=32 | D_K* threshold: D_K* ~ 0.46 x with x = (r m / d_model) log2 m | Hurt |
Closest papers by meaning
- Compositional Attention: Disentangling Search and Retrieval (ICLR 2022 · similarity 0.87)
- Watch Your Step: Learning Node Embeddings via Graph Attention (NeurIPS 2018 · similarity 0.87)
- Bayes optimal learning of attention-indexed models (NeurIPS 2025 · similarity 0.86)
- Virtual Head Attention (NeurIPS 2026 · similarity 0.85)
- Graph Attention is Not Always Beneficial: A Theoretical Analysis of Graph Attention Mechanisms via Contextual Stochastic Block Models (ICML 2025 · similarity 0.85)
- Attention Mechanisms Perspective: Exploring LLM Processing of Graph-Structured Data (ICML 2025 · similarity 0.85)
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.
