Research map / Label-efficient and robust learning / Knowledge distillation
A Comprehensive Overhaul of Feature Distillation
Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, Jin Young Choi
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 |
|---|---|---|---|
| distill at pre-ReLU position instead of end of block (ablation) | feature distillation baseline · CIFAR-100 setup (c) | error rate (%): 26.37 -> 24.81 | Improved |
| use teacher batch norm in training mode (BN student transform) (ablation) | feature distillation baseline · CIFAR-100 setup (c) | error rate (%): 24.81 -> 24.68 | Improved |
| add margin ReLU teacher transform and partial L2 distance (ablation) | feature distillation baseline · CIFAR-100 setup (c) | error rate (%): 24.68 -> 24.08 | Improved |
| teacher batch norm in training mode rather than evaluation mode (ablation) | Proposed · CIFAR-100 setup (c) | error rate (%): 24.54 -> 24.08 | Improved |
| teacher batch norm in training mode rather than evaluation mode (ablation) | FitNets · CIFAR-100 setup (c) | error rate (%): 26.92 -> 26.61 | Improved |
| teacher batch norm in training mode rather than evaluation mode (ablation) | AT · CIFAR-100 setup (c) | error rate (%): 26.42 -> 26.56 | Hurt |
| teacher batch norm in training mode rather than evaluation mode (ablation) | KD · CIFAR-100 setup (c) | error rate (%): 26.45 -> 26.47 | No clear effect |
| move distillation to pre-ReLU position (ablation) | FT · CIFAR-100 setup (c) | error rate (%): 25.91 -> 25.11 | Improved |
| move distillation to pre-ReLU position (ablation) | Jacobian · CIFAR-100 setup (c) | error rate (%): 26.71 -> 26.27 | Improved |
| move distillation to pre-ReLU position (ablation) | FitNets · CIFAR-100 setup (c) | error rate (%): 26.30 -> 26.22 | No clear effect |
| move distillation to pre-ReLU position (ablation) | AT · CIFAR-100 setup (c) | error rate (%): 26.42 -> 26.45 | No clear effect |
| use multiple distillation layers instead of a single layer (ablation) | Proposed · CIFAR-100 setup (e) | error rate (%): 18.17 -> 17.80 | No clear effect |
Closest papers by meaning
- Distilling Object Detectors via Decoupled Features (CVPR 2021 · similarity 0.89)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.89)
- Task-Oriented Feature Distillation (NeurIPS 2020 · similarity 0.89)
- Masked Generative Distillation (ECCV 2022 · similarity 0.89)
- Improve Object Detection with Feature-based Knowledge Distillation: Towards Accurate and Efficient Detectors (ICLR 2021 · similarity 0.88)
- VkD: Improving Knowledge Distillation using Orthogonal Projections (CVPR 2024 · similarity 0.86)
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.
