Research map / Architectures and efficiency / Attention and transformers
MIA-Former: Efficient and Robust Vision Transformers via Multi-Grained Input-Adaptation
Zhongzhi Yu, Yonggan Fu, Sicheng Li, Chaojian Li, Yingyan Lin
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 |
|---|---|---|---|
| Apply multi-grained input-adaptive skipping of blocks, heads, and tokens | DeiT-Small · ImageNet-1K | FLOPs: 20.1% reduction | Improved |
| Apply multi-grained input-adaptive skipping | DeiT-Small · ImageNet-1K under PGD attack | robust accuracy: 2.4% higher | Improved |
| Apply multi-grained input-adaptive skipping | DeiT-Small · ImageNet-1K under FGSM attack | robust accuracy: 3.0% higher | Improved |
| Enable only head-wise input-adaptive skipping instead of all three dimensions (ablation) | MIA-Former on DeiT-Small · ImageNet-1K | accuracy: 79.9 -> 78.7 | Hurt |
| Enable only depth-wise input-adaptive skipping instead of all three dimensions (ablation) | MIA-Former on DeiT-Small · ImageNet-1K | accuracy: 79.9 -> 76.2 | Hurt |
| Enable only token-wise input-adaptive skipping instead of all three dimensions (ablation) | MIA-Former on DeiT-Small · ImageNet-1K | accuracy: 79.9 -> 79.7 | Hurt |
| Use a 75% inherited controller weight initialization rather than 50% (ablation) | MIA-Former RL skipping-policy fine-tuning · ImageNet-1K, target 3.4 GFLOPs | robust accuracy: 21.3 -> 19.8 | Hurt |
| Use a 75% inherited controller weight initialization rather than 50% (ablation) | MIA-Former RL skipping-policy fine-tuning · ImageNet-1K, target 3.4 GFLOPs | clean accuracy: 79.2 -> 79.9 | Improved |
| Use a 100% inherited controller weight initialization rather than 75% (ablation) | MIA-Former RL skipping-policy fine-tuning · ImageNet-1K, target 3.4 GFLOPs | robust accuracy: 19.8 -> 15.7 | Hurt |
| Use a 100% inherited controller weight initialization rather than 75% (ablation) | MIA-Former RL skipping-policy fine-tuning · ImageNet-1K, target 3.4 GFLOPs | clean accuracy: 79.9 -> 79.9 | No clear effect |
| Use a 75% inherited controller weight initialization rather than 50% (ablation) | MIA-Former RL skipping-policy fine-tuning · ImageNet-1K, target 3.9 GFLOPs | robust accuracy: 22.8 -> 20.2 | Hurt |
| Use a 75% inherited controller weight initialization rather than 50% (ablation) | MIA-Former RL skipping-policy fine-tuning · ImageNet-1K, target 3.9 GFLOPs | clean accuracy: 78.3 -> 79.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LeViT-192 — LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference (ICCV 2021) · better on 3 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ReMoE: Region-Mixture Experts for Adversarially-Robust Vision Transformers (CVPR 2026 · similarity 0.87)
- Vision Transformers Are Robust Learners (AAAI 2022 · similarity 0.87)
- Towards Robust Vision Transformers: Path Dependency Analysis and a Simple Two-Stage Adversarial Training (CVPR 2026 · similarity 0.87)
- The Principle of Diversity: Training Stronger Vision Transformers Calls for Reducing All Levels of Redundancy (CVPR 2022 · similarity 0.87)
- Split Adaptation for Pre-trained Vision Transformers (CVPR 2025 · similarity 0.86)
- Vision Transformer Neural Architecture Search for Out-of-Distribution Generalization: Benchmark and Insights (NeurIPS 2024 · 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.
