Research map / Architectures and efficiency / Attention and transformers
The Effect of Attention Head Count on Transformer Approximation
Penghao Yu, Haotian Jiang, Zeyu Bao, Ruoxi Yu, Qianxiao Li
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 |
|---|---|---|---|
| increase number of attention heads to the intrinsic dimension D=4 (ablation) | single-layer transformer · Synthetic sum of 4 maxima (T from 8 to 128) | NMSE | Improved |
| increase sequence length T when heads h < D (ablation) | single-layer transformer · Synthetic sum of 4 maxima | NMSE | Hurt |
| increase sequence length T when heads h >= D (ablation) | single-layer transformer · Synthetic sum of 4 maxima | NMSE | No clear effect |
| increase MLP hidden dimension N (parameter count) with h <= 3 heads (ablation) | single-layer transformer · Synthetic sum of 4 maxima | NMSE | Improved |
| increase MLP hidden dimension N with h = 4 heads (ablation) | single-layer transformer · Synthetic sum of 4 maxima | NMSE | No clear effect |
| increase number of heads above about 12 (ablation) | 2-layer transformer encoder · MS MARCO text retrieval (train accuracy) | accuracy | Improved |
| increase number of heads to about 10 (ablation) | 4-layer ViT · CIFAR-10 (train accuracy, varied sequence length) | accuracy | Improved |
| fix total embedding dimension E=32 instead of fixed per-head dimension (ablation) | single-layer transformer · Synthetic sum of 4 maxima | NMSE | No clear effect |
| use a 2-layer transformer without positional encoding or layer norm (ablation) | transformer · Synthetic sum of 4 maxima | NMSE | No clear effect |
Closest papers by meaning
- Understanding the Expressive Power and Mechanisms of Transformer for Sequence Modeling (NeurIPS 2024 · similarity 0.92)
- Approximation Rate of the Transformer Architecture for Sequence Modeling (NeurIPS 2024 · similarity 0.91)
- Prompting a Pretrained Transformer Can Be a Universal Approximator (ICML 2024 · similarity 0.90)
- Memorization Capacity of Multi-Head Attention in Transformers (ICLR 2024 · similarity 0.89)
- InfoFlow: A Framework for Multi-Layer Transformer Analysis (NeurIPS 2026 · similarity 0.89)
- Are Transformers with One Layer Self-Attention Using Low-Rank Weight Matrices Universal Approximators? (ICLR 2024 · similarity 0.89)
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.
