Research map / Label-efficient and robust learning / Knowledge distillation
Knowledge Distillation from A Stronger Teacher
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace exact KL prediction matching with Pearson correlation over inter-class and intra-class relations | ResNet-18 distillation from ResNet-152 · ImageNet | Top-1 ACC (%): 71.12 -> 72.24 (+1.12) | Helped |
| Replace KL distillation with DIST | Swin-T with Swin-L teacher · ImageNet | Top-1 ACC (%): 81.5 -> 82.3 | Helped |
| Replace KD with DIST | ResNet-18 from ResNet-34 · ImageNet | ACC (%): 71.21 -> 72.07 | Helped |
| Add intra-class correlation to inter-class correlation (ablation) | DIST · ImageNet | ACC (%): 71.63 -> 72.07 | Helped |
| Add inter-class correlation to intra-class correlation (ablation) | DIST · ImageNet | ACC (%): 71.55 -> 72.07 | Helped |
| Replace Pearson correlation with KL divergence for both relations (ablation) | DIST · ImageNet | ACC (%): 72.07 -> 71.62 | Hurt |
| Use only intra-class KL relation rather than inter-class and intra-class Pearson relations (ablation) | DIST · ImageNet | ACC (%): 72.07 -> 70.61 | Hurt |
| Remove classification loss and train with distillation loss only (ablation) | DIST · ImageNet | ACC (%): 72.07 -> 70.65 | Hurt |
| Replace Pearson correlation with cosine similarity (ablation) | DIST without label smoothing · ImageNet | ACC (%): 72.07 -> 71.79 | Hurt |
| Replace Pearson correlation with cosine similarity (ablation) | DIST with label smoothing · ImageNet | ACC (%): 72.18 -> 71.63 | Hurt |
| Add feature mimic loss to DIST (ablation) | Faster R-CNN-R50 · COCO validation, two-stage detectors | AP: 40.4 -> 41.8 | Helped |
| Add feature mimic loss to DIST (ablation) | RetinaNet-R50 · COCO validation, one-stage detectors | AP: 39.8 -> 40.1 | Helped |
Closest papers by meaning
- Knowledge Distillation With the Reused Teacher Classifier (CVPR 2022 · similarity 0.92)
- Better Teacher Better Student: Dynamic Prior Knowledge for Knowledge Distillation (ICLR 2023 · similarity 0.91)
- Less or More From Teacher: Exploiting Trilateral Geometry For Knowledge Distillation (ICLR 2024 · similarity 0.90)
- Distilling Knowledge from Heterogeneous Architectures for Semantic Segmentation (AAAI 2025 · similarity 0.89)
- Understanding the Role of the Projector in Knowledge Distillation (AAAI 2024 · similarity 0.89)
- Wasserstein Distance Rivals Kullback-Leibler Divergence for Knowledge Distillation (NeurIPS 2024 · 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.