Research map / Label-efficient and robust learning / Knowledge distillation
Agree to Disagree: Adaptive Ensemble Knowledge Distillation in Gradient Space
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 averaged teacher loss (AVER) with gradient-space adaptive weighting (AE-KD) | logits-based ensemble KD · CIFAR10, 5 teachers, resnet20 | top-1 accuracy: 91.94 -> 92.50 | Helped |
| replace AVER with AE-KD on a different-architecture student | logits-based ensemble KD · CIFAR10, 5 teachers, MobileNetV2 | top-1 accuracy: 76.11 -> 77.32 | Helped |
| replace AVER with AE-KD | logits-based ensemble KD · CIFAR100, 25 teachers, resnet20 | top-1 accuracy: 71.57 -> 71.84 | Helped |
| replace AVER with AE-KD with teacher and student the same architecture | logits-based ensemble KD · CIFAR10, 25 teachers, resnet56 | top-1 accuracy: 93.73 -> 94.20 | Helped |
| replace AVER with AE-KD for mixed-architecture teachers | logits-based ensemble KD · CIFAR100, resnet20 | top-1 accuracy: 69.40 -> 70.16 | Helped |
| replace AVER with AE-KD at scale | logits-based ensemble KD · ImageNet, 5 teachers | top-1 accuracy: 68.18 -> 69.14 | Helped |
| replace AVER with AE-KD with 3 teachers | logits-based ensemble KD · ImageNet, 3 teachers | top-5 accuracy: 88.39 -> 88.21 | Hurt |
| use feature-based AE-KD* instead of FitNets average | feature-based ensemble KD · CIFAR10, 5 teachers, MobileNetV2 | top-1 accuracy: 78.19 -> 81.16 | Helped |
| use intermediate tolerance C in (1/M, 1) instead of C=1/M (average) or C=1 (ablation) | CIFAR10, ensemble sizes 10 and 20 | top-1 accuracy | Helped |
| increase ensemble size (ablation) | AVER and AE-KD · CIFAR10, resnet20, AVER | top-1 accuracy: 91.94 -> 92.50 | Helped |
Closest papers by meaning
- Single-Teacher View Augmentation: Boosting Knowledge Distillation via Angular Diversity (NeurIPS 2025 · similarity 0.89)
- How to Trade Off the Quantity and Capacity of Teacher Ensemble: Learning Categorical Distribution to Stochastically Employ a Teacher for Distillation (AAAI 2024 · similarity 0.89)
- Ensemble Distribution Distillation via Flow Matching (ICML 2025 · similarity 0.88)
- Toward Student-oriented Teacher Network Training for Knowledge Distillation (ICLR 2024 · similarity 0.87)
- Less or More From Teacher: Exploiting Trilateral Geometry For Knowledge Distillation (ICLR 2024 · similarity 0.86)
- Cooperative Knowledge Distillation: A Learner Agnostic Approach (AAAI 2024 · similarity 0.86)
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.