Research map / Label-efficient and robust learning / Knowledge distillation
AMD: Automatic Multi-step Distillation of Large-scale Vision Models
Cheng Han, Qifan Wang, Sohail A Dianat, Majid Rabbani, Raghuveer Rao, Yi Fang, Qiang Guan, Lifu Huang, Dongfang Liu
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 |
|---|---|---|---|
| use AMD multi-step distillation instead of MMD (manual search over teacher-assistants) | AMD · CIFAR-100 (ViT-Base 10% student) | GPU hours: 20x -> 2.2x | Improved |
| AMD versus manual multi-step distillation in accuracy | AMD · CIFAR-100 (ViT-Tiny 10% student) | top-1 accuracy: 73.12 -> 73.10 | No clear effect |
| add feature-mimicking loss L_feat (ablation) | AMD · CIFAR-100 (ViT-Base 10% student) | top-1 accuracy: 75.24 -> 80.19 | Improved |
| add logit loss L_logit (ablation) | AMD · CIFAR-100 (ViT-Base 10% student) | top-1 accuracy: 78.01 -> 80.19 | Improved |
| add cross-entropy loss L_ce (ablation) | AMD · CIFAR-100 (ViT-Base 10% student) | top-1 accuracy: 78.32 -> 80.19 | Improved |
| add DKD loss on top of AMD objective (ablation) | AMD · CIFAR-100 (ViT-Base 10% student) | top-1 accuracy: 80.19 -> 80.22 | No clear effect |
| add one teacher-assistant versus none (ablation) | AMD · CIFAR-100 (ViT-Base 10% student) | top-1 accuracy: 79.77 -> 80.19 | Improved |
| use more than one teacher-assistant (ablation) | AMD · CIFAR-100 (ViT-Base 10% student) | top-1 accuracy: 80.19 -> 80.24 | No clear effect |
| increase candidate scales m from 1 to 9 (ablation) | AMD · CIFAR-100 (ViT-Base 10% student) | top-1 accuracy: 78.39 -> 80.19 | Improved |
| increase candidate scales m from 9 to 15 (ablation) | AMD · CIFAR-100 (ViT-Base 10% student) | top-1 accuracy: 80.19 -> 80.22 | No clear effect |
| use lambda-NPSD including student term instead of plain NPSD (ablation) | AMD · CIFAR-100 (ViT-Base 10% student) | top-1 accuracy: 80.19 -> 80.14 | No clear effect |
| use AMD instead of ADKD at 5% student scale | AMD · CIFAR-100 (ViT-Base 5% student) | top-1 accuracy: 68.44 -> 75.88 | Improved |
Closest papers by meaning
- Knowledge Distillation: A Good Teacher Is Patient and Consistent (CVPR 2022 · similarity 0.90)
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.90)
- Knowledge Distillation: A Survey (IJCV 2021 · similarity 0.90)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.89)
- DUNE: Distilling a Universal Encoder from Heterogeneous 2D and 3D Teachers (CVPR 2025 · similarity 0.88)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.87)
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.
