Research map / Label-efficient and robust learning / Knowledge distillation
Show, Attend and Distill: Knowledge Distillation via Attention-based Feature Matching
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 |
|---|---|---|---|
| Use attention-based feature links instead of manually selected links | feature distillation · tinyImageNet, ResNet34 → ResNet18 | accuracy: 0.6819 -> 0.6880 | Helped |
| Use attention-based feature links instead of manually selected links | feature distillation · ImageNet, ResNet34 → ResNet18 | accuracy: 0.7137 -> 0.7138 | Helped |
| Use attention-based feature links instead of L2T links | transfer learning · CUB200, MIT67, Stanford40 | accuracy: 0.6505 -> 0.6829; 0.6485 -> 0.6647; 0.6308 -> 0.6792 | Helped |
| Use attention-based feature links instead of L2T links | transfer learning · Stanford Dogs | accuracy: 0.7808 -> 0.7606 | Hurt |
| Use AFD links instead of ordered manually selected links (ablation) | ResNet34 → WRN-28-2 with equal-interval teacher candidates · CIFAR-100 | accuracy: 0.7544 -> 0.7723 | Helped |
| Co-train AFD links rather than use ordered links (ablation) | ResNet34 → WRN-28-2 with equal-interval teacher candidates · CIFAR-100 | accuracy: 0.7544 -> 0.7726 | Helped |
| Use all teacher candidates when student candidates are numerous (ablation) | AFD, ResNet56 → ResNet20 · CIFAR-100, 9 student candidates | accuracy: 0.7115 -> 0.7153 | Helped |
| Use all teacher candidates with only 3 student candidates (ablation) | AFD, ResNet56 → ResNet20 · CIFAR-100, 3 student candidates | accuracy: 0.7140 -> 0.7109 | Hurt |
| Replace L2 distance with L1 distance (ablation) | AFD · CIFAR-100, WRN-40-2 → WRN-16-2 | accuracy: 0.7547 -> 0.7513 | Hurt |
| Replace L2 distance with KL divergence (ablation) | AFD · CIFAR-100, WRN-40-2 → WRN-16-2 | accuracy: 0.7547 -> 0.7523 | Hurt |
| Replace L2 distance with cosine similarity (ablation) | AFD · CIFAR-100, WRN-40-2 → WRN-16-2 | accuracy: 0.7547 -> 0.7541 | Hurt |
| Replace A2 pooling with A1 average pooling (ablation) | AFD · CIFAR-100, WRN-40-2 → WRN-16-2 | accuracy: 0.7547 -> 0.7520 | Hurt |
Closest papers by meaning
- Distilling Knowledge via Knowledge Review (CVPR 2021 · similarity 0.91)
- Cooperative Knowledge Distillation: A Learner Agnostic Approach (AAAI 2024 · similarity 0.89)
- Single-Teacher View Augmentation: Boosting Knowledge Distillation via Angular Diversity (NeurIPS 2025 · similarity 0.89)
- Toward Student-oriented Teacher Network Training for Knowledge Distillation (ICLR 2024 · similarity 0.89)
- Neural Collapse Inspired Knowledge Distillation (AAAI 2025 · similarity 0.89)
- Progressive Distillation Based on Masked Generation Feature Method for Knowledge Graph Completion (AAAI 2024 · 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.