Research map / Label-efficient and robust learning / Knowledge distillation
Markov Knowledge Distillation: Make Nasty Teachers trained by Self-undermining Knowledge Distillation Fully Distillable
En-hui Yang, Linfeng Ye
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 Markov transform (co-learned) to distillation from a nasty teacher (ablation) | MKD-KD · CIFAR-100, nasty R18 teacher, MobileNetV2 student | top-1 accuracy (%): 42.76 -> 71.41 | Improved |
| add power transform on top of Markov transform (ablation) | MKD-KD · CIFAR-100, nasty R18 teacher, MobileNetV2 student | top-1 accuracy (%): 71.37 -> 71.41 | No clear effect |
| use MKD-KD instead of standard KD to distill a nasty teacher | KD · CIFAR-100, nasty R18 teacher, MobileNetV2 student | top-1 accuracy (%): 6.53 -> 71.41 | Improved |
| use MKD-KD instead of Skep., HTC, Avg. for nasty teachers | nasty teacher distillation · CIFAR-10/100, TinyImageNet | top-1 accuracy (%): MKD-KD above LS student in all pairs, up to 3.57% gain | Improved |
| apply Markov transform to normal-teacher distillation (MKD-KD vs KD) | KD · CIFAR-100 and TinyImageNet, normal teachers | top-1 accuracy (%): TinyImageNet R18/SV2: 58.09 -> 60.96; CIFAR-100 R18/MV2: 72.41 -> 72.31 | Mixed |
| apply Markov transform on top of DKD for normal teachers | DKD · CIFAR-100 normal teachers | top-1 accuracy (%): 72.60 -> 73.04 (R18/MV2) | Improved |
| use MKD-KD for cross-domain distillation instead of KD | KD · Cifar-50-50 (nasty R18, MV2) | top-1 accuracy (%): 4.04 -> 79.53 | Improved |
| existing logit-based KD methods applied to nasty teacher | KD, DKD, DIST, NKD, MLD, Skep., HTC, Avg. · CIFAR-10, nasty R18 teacher | top-1 accuracy (%): RC20: LS 92.53, none of the methods exceeds it | Hurt |
Closest papers by meaning
- Undistillable: Making A Nasty Teacher That CANNOT teach students (ICLR 2021 · similarity 0.92)
- Born Again Neural Networks (ICML 2018 · similarity 0.89)
- SGD-Based Knowledge Distillation with Bayesian Teachers: Theory and Guidelines (ICLR 2026 · similarity 0.89)
- Taught Well Learned Ill: Towards Distillation-conditional Backdoor Attack (NeurIPS 2025 · similarity 0.88)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · similarity 0.88)
- Neural Collapse Inspired Knowledge Distillation (AAAI 2025 · 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.
