Research map / Label-efficient and robust learning / Knowledge distillation
REDistill: Robust Estimator Distillation for Balancing Robustness and Efficiency
Ondrej Tybl, Lukas Neumann
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 |
|---|---|---|---|
| replace KL with power divergence (lambda=2/3) in DKD-style loss | DKD · CIFAR-100 (model-agnostic, ResNet32x4->SHN-V2) | top-1 acc: 77.01 -> 77.52 | Improved |
| replace KL with power divergence (lambda=2/3) | DKD · CIFAR-100 (model-agnostic, WRN-40-2->ResNet8x4) | top-1 acc: 75.50 -> 77.21 | Improved |
| add REDistill loss to KD | KD · CIFAR-100 (model-specific, ResNet32x4->SHN-V2) | top-1 acc: 74.45 -> 74.99 | Improved |
| add REDistill loss to MLKD | MLKD · CIFAR-100 (model-specific, ResNet32x4->SHN-V2) | top-1 acc: 78.44 -> 78.82 | Improved |
| add REDistill loss to MLKD | MLKD · CIFAR-100 (model-specific, VGG13->VGG8) | top-1 acc: 75.18 -> 75.24 | Improved |
| use lambda 2/3 instead of 0 (KL) (ablation) | REDistill · CIFAR-100 ResNet32x4->SHN-V2 | top-1 acc: 77.11 -> 77.52 | Improved |
| use lambda 2/3 vs 1 (ablation) | REDistill · CIFAR-100 ResNet32x4->SHN-V2 | top-1 acc: 77.52 vs 77.50 | No clear effect |
| increase lambda to 2 (ablation) | REDistill · CIFAR-100 ResNet32x4->SHN-V2 | top-1 acc: 77.52 -> 76.31 | Hurt |
| use REDistill vs best baseline on VGG13->VGG8 (model-agnostic) | REDistill · CIFAR-100 VGG13->VGG8 | top-1 acc: 74.78 vs LSKD 74.61; RLD 74.40 in model-agnostic, but 74.88 vs RLD 74.93 in model-specific | Mixed |
| REDistill vs RLD on ResNet110->ResNet20 (model-agnostic) | REDistill · CIFAR-100 | top-1 acc: 71.60 vs LSKD 71.67 | Mixed |
Closest papers by meaning
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.90)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 2025 · similarity 0.88)
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025 · similarity 0.88)
- Teacher as a Lenient Expert: Teacher-Agnostic Data-Free Knowledge Distillation (AAAI 2024 · similarity 0.88)
- Can Students Beyond the Teacher? Distilling Knowledge from Teacher’s Bias (AAAI 2025 · similarity 0.88)
- Toward Student-oriented Teacher Network Training for Knowledge Distillation (ICLR 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.
