Research map / Label-efficient and robust learning / Knowledge distillation
Generalized Knowledge Distillation via Relationship Matching
Han-Jia Ye, Su Lu, De‐Chuan Zhan
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 instance-adaptive weights lambda_i (ablation) | ReFilled · CUB GKD (overlap 50%, width 1) | accuracy: 69.14 -> 70.25 | Improved |
| add instance-adaptive weights lambda_i (ablation) | ReFilled · CIFAR-100 GKD (overlap 60%, student (40,2)) | accuracy: 80.02 -> 80.66 | Improved |
| add comparison-matching embedding stage before classifier stage (ablation) | ReFilled · CUB GKD (overlap 50%, width 1) | accuracy: 68.20 -> 68.79 | Improved |
| add LKD instead of plain KD in second stage (ablation) | ReFilled · CIFAR-100 standard KD (student (40,2)) | accuracy: 77.08 -> 77.90 | Improved |
| add LKD instead of plain KD in second stage (ablation) | ReFilled · CIFAR-100 standard KD (student (16,1)) | accuracy: 67.14 -> 68.13 | Improved |
| add comparison matching (CM) in first stage (ablation) | ReFilled · CIFAR-100 NCM accuracy (student (40,2)) | NCM accuracy: 55.47 -> 62.12 | Improved |
| two-stage training instead of one-stage with balance gamma (ablation) | ReFilled · CUB GKD (overlap 50%, width 1) | accuracy: 74.25 -> 75.13 | Improved |
| use PL-weight rather than Gap-weight (ablation) | ReFilled · CUB GKD (overlap 50%, width 1) | accuracy: 70.13 -> 70.25 | Improved |
| use Gap-weight over no adaptive weights (ablation) | ReFilled · CIFAR-100 GKD (overlap 60%, student (16,1)) | accuracy: 68.72 -> 69.61 | Improved |
| apply ReFilled when teacher on CUB and target is Stanford Dogs | ReFilled · Stanford Dogs (width 1) | accuracy: 72.35 -> 72.27 | No clear effect |
| ReFilled vs RKD in standard KD | ReFilled · CUB standard KD (width 1) | accuracy: 77.72 -> 79.33 | Improved |
| ReFilled vs best baseline SSKD at student (40,1) | ReFilled · CIFAR-100 standard KD (student (40,1)) | accuracy: 72.71 -> 72.54 | Hurt |
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.
- SEED — SEED: Self-supervised Distillation For Visual Representation (ICLR 2021) · better on 2 of 3 shared results
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
- Distilling Cross-Task Knowledge via Relationship Matching (CVPR 2020 · similarity 0.95)
- Student Customized Knowledge Distillation: Bridging the Gap Between Student and Teacher (ICCV 2021 · similarity 0.90)
- Similarity-Preserving Knowledge Distillation (ICCV 2019 · similarity 0.90)
- Knowledge Distillation from A Stronger Teacher (NeurIPS 2022 · similarity 0.90)
- Densely Guided Knowledge Distillation Using Multiple Teacher Assistants (ICCV 2021 · similarity 0.90)
- Neural Collapse Inspired Knowledge Distillation (AAAI 2025 · 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.
