Research map / Label-efficient and robust learning / Knowledge distillation
Automated Knowledge Distillation via Monte Carlo Tree Search
Lujun Li, Peijie Dong, Zimian Wei, Ya Yang
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 |
|---|---|---|---|
| use MCTS instead of random search for distiller search (ablation) | Auto-KD · CIFAR-100 ResNet20 | top-1 accuracy (%): 71.55 -> 72.52 | Improved |
| use the tree-structured search space instead of a naive organization (ablation) | Auto-KD · CIFAR-100 ResNet20 | top-1 accuracy (%): 70.32 -> 72.52 (MCTS) | Improved |
| add extended search space operations (ablation) | Auto-KD · CIFAR-100 ResNet20 | top-1 accuracy (%): 72.52 -> 72.65 | Improved |
| add CKA representation gap to the loss-based reward (ablation) | Auto-KD · CIFAR-100 ResNet20 | top-1 accuracy (%): 72.36 -> 72.52 | Improved |
| use loss value as reward instead of accuracy (ablation) | Auto-KD · CIFAR-100 ResNet20 | top-1 accuracy (%): 72.25 -> 72.36 | Improved |
| offline teacher feature storage during search (ablation) | Auto-KD search · CIFAR-100 ResNet20 | search time / final accuracy: 6,036 s -> 4,654 s; Acc-2 72.65 -> 72.65 | Improved |
| add subsets-20% and early-stop-20% proxy settings (ablation) | Auto-KD search · CIFAR-100 ResNet20 | final accuracy Acc-2 / time: Acc-2 72.61 -> 72.52; time 3,816 s -> 150.6 s | Mixed |
| add 50% sparse student training in search (ablation) | Auto-KD search · CIFAR-100 ResNet20 | final accuracy Acc-2: 72.65 -> 72.61 | No clear effect |
| use train-free factors with Auto-KD for multi-layer distillation | Auto-KD · CIFAR-100 multi-layer | top-1 accuracy (%): 71.93 -> 72.55 (vs KR+DistPro) | Improved |
| apply Auto-KD on object detection with distillers searched on classification | Faster RCNN-R50 / RetinaNet-R50 · MS-COCO | AP: 38.40 -> 42.40; 37.40 -> 41.10 | Improved |
| Auto-KD vs MGD on RetinaNet detection | RetinaNet-R50 · MS-COCO | AP: 41.00 -> 41.10 | No clear effect |
Closest papers by meaning
- KD-Zero: Evolving Knowledge Distiller for Any Teacher-Student Pairs (NeurIPS 2023 · similarity 0.91)
- DistPro: Searching a Fast Knowledge Distillation Process via Meta Optimization (ECCV 2022 · similarity 0.91)
- DTO-KD: Dynamic Trade-off Optimization for Effective Knowledge Distillation (ICLR 2026 · similarity 0.90)
- Search to Distill: Pearls Are Everywhere but Not the Eyes (CVPR 2020 · similarity 0.90)
- Knowledge Distillation and Student-Teacher Learning for Visual Intelligence: A Review and New Outlooks (TPAMI 2021 · similarity 0.90)
- SGD-Based Knowledge Distillation with Bayesian Teachers: Theory and Guidelines (ICLR 2026 · 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.
