Research map / Label-efficient and robust learning / Knowledge distillation
Good Teachers Explain: Explanation-Enhanced Knowledge Distillation
Amin Parchami-Araghi, Moritz Böhle, Sukrut Rao, Bernt Schiele
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 explanation-matching loss (e2KD, GradCAM) to logit KD, 50 shots | KD, ResNet-34 -> ResNet-18 · ImageNet val, 50 shots | top-1 accuracy (%): 49.8 -> 54.9 | Improved |
| add e2KD to KD with full data | KD, ResNet-34 -> ResNet-18 · ImageNet val, full data | top-1 accuracy (%): 71.8 -> 71.8 | No clear effect |
| add e2KD (B-cos) to KD | KD, B-cos ResNet-34 -> B-cos ResNet-18 · ImageNet val, 50 shots | teacher agreement (%): 38.4 -> 48.4 | Improved |
| add e2KD to CRD | CRD · ImageNet val, 200 shots | top-1 accuracy (%): 51.0 -> 54.1 | Improved |
| add e2KD across different student architectures | KD on Waterbirds-100 · Waterbirds-100 out-of-distribution | accuracy (%): ConvNeXt 20.5 -> 32.2; ShuffleNet 23.1 -> 37.0 | Improved |
| add e2KD so the student matches a guided teacher's explanations | KD on PASCAL VOC · PASCAL VOC EPG teacher | F1: 60.1 -> 67.6 | Improved |
| distill CNN teacher to ViT student with e2KD | B-cos ViT-Tiny student · ImageNet val | teacher agreement (%): 70.1 -> 71.8 | Improved |
| use frozen pre-computed explanations with e2KD instead of frozen-logit KD (ablation) | B-cos DenseNet-169 -> ResNet-18, 50 shots · ImageNet val | top-1 accuracy (%): 33.4 -> 38.7 | Improved |
| remove 2x2 down-sampling of CAMs in CAT-KD without rescaling the loss (ablation) | CAT-KD · ImageNet 50 shots | top-1 accuracy (%): 32.2 -> 24.4 | Hurt |
| replace CE with KL-divergence and use top-1 explanation in CAT-KD (ablation) | CAT-KD · ImageNet 50 shots | top-1 accuracy (%): 35.0 -> 53.1 | Improved |
| replace CAM with GradCAM in the CAT-KD variant (ablation) | CAT-KD variant · ImageNet 50 shots | top-1 accuracy (%): 53.1 -> 52.9 | No clear effect |
| longer teaching schedule and mixup for KD (ablation) | KD on Waterbirds-100 · Waterbirds-100 | agreement | Improved |
Closest papers by meaning
- Can Students Beyond the Teacher? Distilling Knowledge from Teacher’s Bias (AAAI 2025 · similarity 0.91)
- A Good Teacher Adapts Their Knowledge for Distillation (ICCV 2025 · similarity 0.91)
- Learning Student-Friendly Teacher Networks for Knowledge Distillation (NeurIPS 2021 · similarity 0.90)
- Knowledge Condensation Distillation (ECCV 2022 · similarity 0.90)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 2025 · similarity 0.89)
- Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling (ICLR 2025 · similarity 0.89)
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.
