Research map / Label-efficient and robust learning / Knowledge distillation
Generalization Matters: Loss Minima Flattening via Parameter Hybridization for Efficient Online Knowledge Distillation
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 parameter hybridization with HWM and regular fusion | OKDPH · CIFAR-10 | top-1 accuracy (%): 95.01 vs 94.32 | Helped |
| Use parameter hybridization with HWM and regular fusion | OKDPH · CIFAR-100 | top-1 accuracy (%): 79.68 vs 78.37 | Helped |
| Add KD loss to CE-only training (ablation) | OKDPH · CIFAR-10 | accuracy (%): +1.21 | Helped |
| Add HWM-student parameter fusion (ablation) | OKDPH · CIFAR-10 | accuracy (%): +0.28 | Helped |
| Add HWM classification loss (ablation) | OKDPH · CIFAR-10 | accuracy (%): +0.26 | Helped |
| Add KD loss to CE-only training (ablation) | OKDPH · CIFAR-100 | accuracy (%): +0.55 | Helped |
| Add HWM-student parameter fusion (ablation) | OKDPH · CIFAR-100 | accuracy (%): +0.29 | Helped |
| Add HWM classification loss (ablation) | OKDPH · CIFAR-100 | accuracy (%): +0.50 | Helped |
| Set fusion ratio γ to 0.0 instead of the selected 0.5 (ablation) | OKDPH · CIFAR-10 | accuracy (%): 95.01 -> 94.17 | Hurt |
| Set fusion interval Δ to 5e instead of the selected 1e (ablation) | OKDPH · CIFAR-10 | accuracy (%): 95.01 -> 93.79 | Hurt |
| Train with parameter hybridization | OKDPH · CIFAR-10, noisy and limited training data | top-1 accuracy (%) | Helped |
Closest papers by meaning
- DA-KD: Difficulty-Aware Knowledge Distillation for Efficient Large Language Models (ICML 2025 · similarity 0.88)
- Single-Teacher View Augmentation: Boosting Knowledge Distillation via Angular Diversity (NeurIPS 2025 · similarity 0.88)
- Over-parameterized Student Model via Tensor Decomposition Boosted Knowledge Distillation (NeurIPS 2024 · similarity 0.88)
- MiniPLM: Knowledge Distillation for Pre-training Language Models (ICLR 2025 · similarity 0.88)
- DDK: Distilling Domain Knowledge for Efficient Large Language Models (NeurIPS 2024 · similarity 0.87)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.87)
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.