Research map / Architectures and efficiency / Attention and transformers
Is Attention Better Than Matrix Decomposition?
Zhengyang Geng, Meng-Hao Guo, Hongxu Chen, Xia Li, Ke Wei, Zhouchen Lin
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 |
|---|---|---|---|
| Replace self-attention with Hamburger | HamNet · PASCAL VOC test, w/o COCO pretraining | mIoU(%): 84.4 -> 85.9 | Improved |
| Replace self-attention with Hamburger | HamNet · PASCAL Context Val | mIoU(%): 54.8 -> 55.2 | Improved |
| Replace self-attention with Hamburger | HamGAN · ImageNet 128×128 | FID: 18.28 -> 16.05 | Improved |
| Replace YLG with Hamburger | HamGAN-strong · ImageNet 128×128 | FID: 15.94 -> 14.77 | Improved |
| Add matrix decomposition ham (ablation) | Hamburger with NMF · PASCAL VOC val | mIoU(%): 75.8(75.6) -> 78.3(77.8) | Improved |
| Add upper bread (ablation) | Hamburger with NMF · PASCAL VOC val | mIoU(%): 77.0(76.8) -> 78.3(77.8) | Improved |
| Add lower bread (ablation) | Hamburger with NMF · PASCAL VOC val | mIoU(%): 77.3(77.2) -> 78.3(77.8) | Improved |
| Add both linear bread transformations around ham (ablation) | Hamburger with NMF · PASCAL VOC val | mIoU(%): 77.0(76.8) -> 78.3(77.8) | Improved |
| Use one-step gradient instead of BPTT (ablation) | VQ ham · PASCAL VOC val | mIoU: 76.6(76.3) -> 77.7(77.4) | Improved |
| Use one-step gradient instead of BPTT (ablation) | CD ham · PASCAL VOC val | mIoU: 75.0(74.6) -> 78.1(77.5) | Improved |
| Use one-step gradient instead of BPTT (ablation) | NMF ham · PASCAL VOC val | mIoU: 77.4(77.0) -> 78.3(77.8) | Improved |
| Use random rather than learned dictionary initialization (ablation) | NMF ham · PASCAL VOC val | mIoU: 76.8(76.5) -> 78.3(77.8) | Improved |
Closest papers by meaning
- Bayes optimal learning of attention-indexed models (NeurIPS 2025 · similarity 0.87)
- Self-attention Networks Localize When QK-eigenspectrum Concentrates (ICML 2024 · similarity 0.85)
- Scale-invariant attention (NeurIPS 2025 · similarity 0.85)
- Customizing the Inductive Biases of Softmax Attention using Structured Matrices (ICML 2025 · similarity 0.85)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.85)
- Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers (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.
