Research map / Label-efficient and robust learning / Knowledge distillation
Improving Knowledge Distillation via Category Structure
Zailiang Chen, Xianxian Zheng, Hailan Shen, Ziyang Zeng, Yukun Zhou, Rongchang Zhao
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 intra-category structure loss to KD (ablation) | KD student · Tiny ImageNet | top-1 accuracy: 49.53 -> 52.51 | Improved |
| add inter-category structure loss to KD (ablation) | KD student · Tiny ImageNet | top-1 accuracy: 49.53 -> 52.48 | Improved |
| add inter-category loss on top of intra-category loss (ablation) | CSKD · Tiny ImageNet | top-1 accuracy: 52.51 -> 53.66 | Improved |
| add intra-category loss on top of inter-category loss (ablation) | CSKD · Tiny ImageNet | top-1 accuracy: 52.48 -> 53.66 | Improved |
| add inter-category loss on top of intra-category loss (ablation) | CSKD · Tiny ImageNet | top-5 accuracy: 78.36 -> 78.75 | Improved |
| add both category structure losses to KD (ablation) | KD student · Tiny ImageNet | top-5 accuracy: 74.90 -> 78.75 | Improved |
| use CSKD instead of CCKD | student ResNet18 0.25 · Tiny ImageNet | top-1 accuracy: 53.14 -> 53.66 | Improved |
| use CSKD instead of KD | student ResNet18 0.25 · CIFAR-10 | accuracy: average 1.52% improvement | Improved |
| use CSKD instead of CE-only training | student networks · CIFAR-100 | accuracy: 2.82%~4.15% | Improved |
| distill with CSKD from R101 0.5 teacher into R50 0.25 | student ResNet50 0.25 · CIFAR-10 | accuracy: 92.91 -> 93.20 | Improved |
| distill with CSKD from R152 0.5 teacher into R50 0.25 | student ResNet50 0.25 · CIFAR-100 | accuracy: 72.15 -> 72.60 | Improved |
| use CSKD instead of KDGAN (R101 0.5/R18 0.25) | student ResNet18 0.25 · CIFAR-10 | accuracy: 91.54 -> 91.99 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- KDGAN — KDGAN: Knowledge Distillation with Generative Adversarial Networks (NeurIPS 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Relational Knowledge Distillation (CVPR 2019 · similarity 0.89)
- Complementary Relation Contrastive Distillation (CVPR 2021 · similarity 0.89)
- Cooperative Knowledge Distillation: A Learner Agnostic Approach (AAAI 2024 · similarity 0.88)
- Knowledge Distillation from Internal Representations (AAAI 2020 · similarity 0.88)
- Knowledge Distillation from A Stronger Teacher (NeurIPS 2022 · similarity 0.87)
- Local Dense Logit Relations for Enhanced Knowledge Distillation (ICCV 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.
