Research map / Architectures and efficiency / Attention and transformers
Mitigating Over-smoothing in Transformers via Regularized Nonlocal Functionals
Tam Minh Nguyen, Tan Minh Nguyen, Richard Baraniuk
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 the fidelity regularizer to self-attention | DeiT-Tiny · ImageNet | Top-1 accuracy: 72.17 -> 73.01 | Improved |
| Add the fidelity regularizer to self-attention | DeiT-Tiny · ADE20K | SS MIoU: 35.72 -> 37.24 | Improved |
| Add the fidelity regularizer to self-attention | Softmax Transformer · WikiText-103 validation | Valid PPL: 33.15 -> 32.60 | Improved |
| Combine NeuTRENO with FeatScale | DeiT-Tiny · ImageNet | Top-1 accuracy: 72.346 -> 73.23 | Improved |
| Set λ̃ to 0.1 (ablation) | NeuTRENO DeiT · ADE20K | SS MIoU: 35.72 -> 34.94 | Hurt |
| Set λ̃ to 0.2 (ablation) | NeuTRENO DeiT · ADE20K | SS MIoU: 35.72 -> 35.60 | Hurt |
| Set λ̃ to 0.4 (ablation) | NeuTRENO DeiT · ADE20K | SS MIoU: 35.72 -> 37.54 | Improved |
| Set λ̃ to 0.5 (ablation) | NeuTRENO DeiT · ADE20K | SS MIoU: 35.72 -> 37.38 | Improved |
| Set λ̃ to 0.6 (ablation) | NeuTRENO DeiT · ADE20K | SS MIoU: 35.72 -> 37.24 | Improved |
| Set λ̃ to 0.8 (ablation) | NeuTRENO DeiT · ADE20K | SS MIoU: 35.72 -> 36.71 | Improved |
| Set λ̃ to 1.0 (ablation) | NeuTRENO DeiT · ADE20K | SS MIoU: 35.72 -> 36.37 | Improved |
| Set λ̃ to 2.0 (ablation) | NeuTRENO DeiT · ADE20K | SS MIoU: 35.72 -> 26.25 | Hurt |
Closest papers by meaning
- Attention-Only Transformers via Unrolled Subspace Denoising (ICML 2025 · similarity 0.88)
- Viewing Transformers Through the Lens of Long Convolutions Layers (ICML 2024 · similarity 0.88)
- Wavy Transformer (NeurIPS 2025 · similarity 0.87)
- Demystifying Singular Defects in Large Language Models (ICML 2025 · similarity 0.87)
- Transformer Meets Twicing: Harnessing Unattended Residual Information (ICLR 2025 · similarity 0.86)
- Transformer Layers as Painters (AAAI 2025 · 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.
