Research map / Label-efficient and robust learning / Knowledge distillation
Direction Sensitivity–Based Knowledge Distillation: Optimization-Aware Low-Rank Knowledge Transfer
Yongkai Liao, Xinxing Chen, Zhongzheng Fu, Haoyuan Wang, Jian Huang
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 DSKD direction-sensitive selection | BERT-of-Theseus · GLUE (average) | average score: 77.7 -> 79.2 | Improved |
| add DSKD | LGTM · GLUE (average) | average score: 79.3 -> 79.8 | Improved |
| add DSKD | DBKD · GLUE (average) | average score: 72.5 -> 73.7 | Improved |
| add DSKD | AD-KD · GLUE (average) | average score: 79.8 -> 80.4 | Improved |
| add DSKD | AD-KD · GLUE RTE | accuracy: 67.5 -> 66.9 | Hurt |
| use sensitivity-based directions instead of top singular values (SVD) (ablation) | BERT-of-Theseus · GLUE RTE | accuracy: 65.7 -> 67.6 | Improved |
| select directions randomly instead of by sensitivity (ablation) | BERT-of-Theseus · GLUE RTE | accuracy: 67.6 -> 63.5 | Hurt |
| combine first- and second-order sensitivity instead of first-order only (ablation) | DSKD · GLUE RTE | accuracy: 67.2 -> 67.6 | Improved |
| combine first- and second-order sensitivity instead of second-order only (ablation) | DSKD · GLUE RTE | accuracy: 66.8 -> 67.6 | Improved |
| use very large distillation weight lambda=10 (ablation) | DSKD · GLUE RTE | accuracy: 67.6 -> 64.7 | Hurt |
| tune lambda between 0 and 1 versus lambda=0 (ablation) | DSKD · GLUE CoLA | accuracy: +5.8% on CoLA | Improved |
| update directions every Ts=100 steps instead of Ts=1 (ablation) | DSKD · GLUE RTE | accuracy: 64.7 -> 67.6 | Improved |
Closest papers by meaning
- Adaptive Dual Guidance Knowledge Distillation (AAAI 2025 · similarity 0.91)
- DistPro: Searching a Fast Knowledge Distillation Process via Meta Optimization (ECCV 2022 · similarity 0.91)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · similarity 0.90)
- Neural Collapse Inspired Knowledge Distillation (AAAI 2025 · similarity 0.89)
- Knowledge Distillation with Auxiliary Variable (ICML 2024 · similarity 0.89)
- DA-KD: Difficulty-Aware Knowledge Distillation for Efficient Large Language Models (ICML 2025 · 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.
