Research map / Label-efficient and robust learning / Knowledge distillation
SGD-Based Knowledge Distillation with Bayesian Teachers: Theory and Guidelines
Itai Morad, Nir Shlezinger, Yonina C. Eldar
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 |
|---|---|---|---|
| use a Bayesian (VI) teacher instead of a deterministic teacher | KD student · CIFAR-100 ResNet-50 -> WRN-16-2 | student accuracy (%): 69.36 -> 73.63 (+4.27) | Improved |
| use a Bayesian teacher instead of deterministic | KD student · CIFAR-100 VGG-13 -> WRN-40-1 | student accuracy (%): 67.90 -> 71.37 (+3.47) | Improved |
| use a Bayesian teacher instead of deterministic | KD student · CIFAR-100 WRN-40-2 -> WRN-16-2 | student accuracy (%): 70.80 -> 72.94 (+2.14) | Improved |
| use a Bayesian teacher instead of deterministic | KD student · CIFAR-100 ResNet-18 -> ResNet-18 | student accuracy (%): 75.92 -> 76.92 (+1.00) | Improved |
| convert deterministic teacher to Bayesian via Laplace approximation | KD student · CIFAR-100 ResNet-50 -> WRN-16-2 | student accuracy (%): 69.36 -> 72.52 (+3.16) | Improved |
| convert deterministic teacher via Laplace approximation | KD student · CIFAR-100 ResNet-50 -> ResNet-34 | student accuracy (%): 76.82 -> 76.23 (-0.59) | Hurt |
| Laplace approximation teacher | teacher accuracy · CIFAR-100 ResNet-50 teacher | teacher accuracy (%): 75.41 -> 74.83 (-0.58) | Hurt |
| train teacher with MSE loss instead of CE | KD student · CIFAR-100 ResNet-50 -> WRN-16-2 | student accuracy (%): 69.36 -> 68.55 (-0.81) | Hurt |
| fine-tune teacher with MCMI | KD student · CIFAR-100 ResNet-50 -> WRN-16-2 | student accuracy (%): 69.36 -> 69.58 (+0.22) | No clear effect |
| test-time data augmentation teacher (TTDA) | KD student · CIFAR-100 ResNet-50 -> WRN-16-2 | student accuracy (%): 69.36 -> 69.24 (-0.12) | No clear effect |
| distill from Bayesian teacher for stability | KD student · CIFAR-100 ResNet-50 -> ResNet-18 | convergence noise: 30% less noise | Improved |
| supervise with true class probabilities instead of one-hot labels | SGD student · synthetic experiment | loss neighborhood | Improved |
Closest papers by meaning
- Bayesian Knowledge Distillation: A Bayesian Perspective of Distillation with Uncertainty Quantification (ICML 2024 · similarity 0.93)
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025 · similarity 0.91)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 2025 · similarity 0.91)
- Can Students Beyond the Teacher? Distilling Knowledge from Teacher’s Bias (AAAI 2025 · similarity 0.90)
- Toward Student-oriented Teacher Network Training for Knowledge Distillation (ICLR 2024 · similarity 0.90)
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.90)
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.
