Research map / Label-efficient and robust learning / Knowledge distillation
Task-Oriented Feature Distillation
Linfeng Zhang, Yukang Shi, Zuoqiang Shi, Kaisheng Ma, Chenglong Bao
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 auxiliary classifiers (task loss) to logit+feature distillation (ablation) | TOFD · CIFAR100 ResNet18 | top-1 accuracy (%): 78.81 -> 82.31 | Improved |
| train with auxiliary classifiers only (multi-exit) (ablation) | ResNet18 baseline · CIFAR100 | top-1 accuracy (%): 77.09 -> 79.99 | Improved |
| add orthogonal loss on feature resizing layer (ablation) | TOFD · CIFAR100 ResNet18 | top-1 accuracy (%): 82.31 -> 82.92 | Improved |
| add logit distillation to feature distillation (ablation) | TOFD · CIFAR100 ResNet18 | top-1 accuracy (%): 78.57 -> 78.81 | Improved |
| vary alpha from 0.01 to 0.07 (ablation) | TOFD · CIFAR100 ResNet18 | top-1 accuracy (%): 82.55-82.90 range | No clear effect |
| vary beta from 0.2 to 0.8 (ablation) | TOFD · CIFAR100 ResNet18 | top-1 accuracy (%): 82.51 -> 82.92 | No clear effect |
| add a fourth auxiliary classifier (ablation) | TOFD · CIFAR100 ResNet18 | top-1 accuracy (%) | No clear effect |
| TOFD vs vanilla KD | TOFD · CIFAR100 ResNet18 | top-1 accuracy (%): 78.34 -> 82.92 | Improved |
| TOFD vs baseline training on ImageNet | TOFD · ImageNet ResNet18 | top-1 accuracy (%): 69.76 -> 70.92 | Improved |
| TOFD vs baseline on 3D classification | TOFD · ModelNet10 ResGCN8 | top-1 accuracy (%): 92.73 -> 94.38 | Improved |
| KD, FitNet, SD vs baseline on 3D classification | KD, FitNet, SD · ModelNet10/40 | top-1 accuracy (%) | Mixed |
| KD vs baseline on CIFAR10 PreactResNet18 | KD · CIFAR10 PreactResNet18 | top-1 accuracy (%): 94.20 -> 93.74 | Hurt |
Closest papers by meaning
- A Comprehensive Overhaul of Feature Distillation (ICCV 2019 · similarity 0.89)
- Improved Feature Distillation via Projector Ensemble (NeurIPS 2022 · similarity 0.88)
- Decoupled Knowledge Distillation (CVPR 2022 · similarity 0.88)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.88)
- Feature Normalized Knowledge Distillation for Image Classification (ECCV 2020 · similarity 0.88)
- VkD: Improving Knowledge Distillation using Orthogonal Projections (CVPR 2024 · similarity 0.88)
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.
