Research map / Label-efficient and robust learning / Knowledge distillation
Generative Diffusion Prior Distillation for Long-Context Knowledge Transfer
Hewa Dehigahawattage Nilushika Udayangani, Nandakishor Desai, Marimuthu Palaniswami
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 |
|---|---|---|---|
| train diffusion prior with warm-up phase (E_warm=300) instead of none (ablation) | GDPD · StarLightCurves e=0.5L | AUC-PRC: 85.56 -> 97.26 | Improved |
| jointly train diffusion prior with GDPD loss without warm-up (ablation) | GDPD · StarLightCurves e=0.5L | AUC-PRC: 85.56 -> 89.72 | Improved |
| train the diffusion prior at all (ablation) | GDPD · StarLightCurves e=0.5L | AUC-PRC: 83.12 -> 97.26 | Improved |
| combine task loss with GDPD loss instead of GDPD only (ablation) | GDPD · StarLightCurves e=0.5L | AUC-PRC: 85.72 -> 97.26 | Improved |
| use GDPD loss instead of logit-KD loss (ablation) | GDPD · StarLightCurves e=0.5L | AUC-PRC: 94.51 -> 97.26 | Improved |
| use GDPD loss instead of feature-KD loss (ablation) | GDPD · StarLightCurves e=0.5L | AUC-PRC: 83.81 -> 97.26 | Improved |
| use more posterior samples J=4 instead of J=1 (ablation) | GDPD · StarLightCurves e=0.5L | AUC-PRC: 97.26 -> 97.87 | No clear effect |
| use GDPD instead of Base-KD for partial classification | LSTM3-100 · 12 UCR e=0.5L | avg AUC-PRC: 78.14 -> 84.64 | Improved |
| use GDPD instead of feature KD (Fits) for compression to LSTM1-8 | LSTM3-100 · 12 UCR e=0.5L | avg AUC-PRC: 59.84 -> 72.76 | Improved |
| use GDPD for self-distillation instead of Base-KD | LSTM3-100 · 12 UCR e=L | avg AUC-PRC: 89.29 -> 91.21 | Improved |
| use GDPD for self-distillation instead of Base-KD | ResNet32-64 · 12 UCR e=L | avg AUC-PRC: 98.52 -> 98.58 | No clear effect |
Closest papers by meaning
- Progressively Knowledge Distillation via Re-parameterizing Diffusion Reverse Process (AAAI 2024 · similarity 0.89)
- CURE: Context-driven Diffusion with Progressive Expansion for Single Domain Generalization in Time Series Classification (ICML 2026 · similarity 0.89)
- Improving Knowledge Distillation via Regularizing Feature Direction and Norm (ECCV 2024 · similarity 0.88)
- Towards Universal Dataset Distillation via Task-Driven Diffusion (CVPR 2025 · similarity 0.86)
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.86)
- Weight Diffusion for Future: Learn to Generalize in Non-Stationary Environments (NeurIPS 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.
