Research map / Label-efficient and robust learning / Knowledge distillation
DOT: A Distillation-Oriented Trainer
Borui Zhao, Quan Cui, Renjie Song, Jiajun Liang
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 larger momentum to distillation gradient than task gradient (DOT) | KD · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy (%): 73.33 -> 75.12 | Improved |
| apply DOT to KD on a harder dataset | KD · Tiny-ImageNet (ResNet18 -> MobileNetV2) | top-1 accuracy (%): 58.35 -> 64.01 | Improved |
| apply DOT to KD on ImageNet-1k | KD · ImageNet-1k (ResNet50 -> MobileNetV1) | top-1 accuracy (%): 70.50 -> 73.09 | Improved |
| apply DOT to DKD | DKD · ImageNet-1k (ResNet50 -> MobileNetV1) | top-1 accuracy (%): 72.05 -> 73.33 | Improved |
| apply DOT to feature-based CRD | CRD · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy (%): 75.51 -> 75.99 | Improved |
| use negative Delta so task loss has larger momentum (ablation) | DOT · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy (%): 73.33 -> 61.58 | Hurt |
| use positive Delta between 0.025 and 0.09 (ablation) | KD · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy (%): 73.33 -> 75.12 | Improved |
| tune the SGD momentum mu without DOT (ablation) | KD · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy (%): 73.33 -> 73.66 | No clear effect |
| adjust the task loss weight alpha (ablation) | KD · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy (%): 73.33 -> 73.56 | No clear effect |
| adjust the task loss weight alpha (ablation) | KD · Tiny-ImageNet (ResNet18 -> MobileNetV2) | top-1 accuracy (%): 58.35 -> 59.23 | No clear effect |
| train KD for longer epochs (240 to 300) (ablation) | KD · CIFAR-100 (ResNet32x4 -> ResNet8x4) | training task loss: 0.3844 -> 0.3818 | No clear effect |
| apply DOT to DKD on CIFAR-100 ShuffleNetV2 student, top-5 | DKD · CIFAR-100 (ResNet32x4 -> ShuffleNetV2) | top-5 accuracy (%): (-0.06) | Hurt |
Closest papers by meaning
- Shadow Knowledge Distillation: Bridging Offline and Online Knowledge Transfer (NeurIPS 2022 · similarity 0.89)
- Enhancing Logits Distillation with Plug&Play Kendall's $\tau$ Ranking Loss (ICML 2025 · similarity 0.89)
- Understanding the Role of the Projector in Knowledge Distillation (AAAI 2024 · similarity 0.88)
- Progressively Knowledge Distillation via Re-parameterizing Diffusion Reverse Process (AAAI 2024 · similarity 0.88)
- Gap Preserving Distillation by Building Bidirectional Mappings with A Dynamic Teacher (ICLR 2025 · similarity 0.88)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · 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.
