Research map / Label-efficient and robust learning / Knowledge distillation
MetaDistiller: Network Self-Boosting via Meta-Learned Top-Down Distillation
Benlin Liu, Yongming Rao, Jiwen Lu, Jie Zhou, Cho-Jui Hsieh
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 meta-learned top-down soft targets for intermediate outputs | ResNet18 self-boosting · CIFAR-100 | top-1 accuracy: 78.25 -> 79.05 | Improved |
| add meta-learned top-down soft targets | ResNet50 · CIFAR-100 | top-1 accuracy: 79.71 -> 80.85 | Improved |
| add meta-learned soft targets | VGG19(BN) · CIFAR-100 | top-1 accuracy: 74.27 -> 75.12 | Improved |
| add meta-learned soft targets | ResNet18 · ImageNet | top-1 accuracy: 69.84 -> 70.52 | Improved |
| add meta-learned soft targets | VGG19(BN) · ImageNet | top-1 accuracy: 72.45 -> 73.40 | Improved |
| add meta-learned soft targets | ResNet50 · ImageNet | top-1 accuracy: 75.24 -> 76.11 | Improved |
| use MetaDistiller instead of a ResNet152 teacher with FitNet | ResNet50 student · CIFAR-100 | top-1 accuracy: 80.20 -> 80.85 | Improved |
| use MetaDistiller instead of a ResNet152 teacher with AT | ResNet18 student · CIFAR-100 | top-1 accuracy: 78.81 -> 79.05 | Improved |
| ensemble all intermediate outputs at inference | MetaDistiller ResNet18 · CIFAR-100 | top-1 accuracy: 79.05 -> 80.05 | Improved |
| add generated soft targets on top of hard-label supervision (vs DSN) (ablation) | ResNet18 · CIFAR-100 | top-1 accuracy: 78.01 -> 79.05 | Improved |
| add generated soft targets on top of hard-label supervision (vs DSN) (ablation) | ResNet50 · CIFAR-100 | top-1 accuracy: 79.33 -> 80.85 | Improved |
| add generated soft targets on top of hard-label supervision (vs DSN) (ablation) | ResNet101 · CIFAR-100 | top-1 accuracy: 80.54 -> 81.39 | Improved |
Closest papers by meaning
- DistPro: Searching a Fast Knowledge Distillation Process via Meta Optimization (ECCV 2022 · similarity 0.87)
- Self-Distillation From the Last Mini-Batch for Consistency Regularization (CVPR 2022 · similarity 0.86)
- Self-Distillation: Towards Efficient and Compact Neural Networks (TPAMI 2021 · similarity 0.85)
- Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks (ICLR 2025 · similarity 0.84)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.84)
- Hybrid Data-Free Knowledge Distillation (AAAI 2025 · similarity 0.83)
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.
