Research map / Label-efficient and robust learning / Knowledge distillation
Knowledge Transfer with Jacobian Matching
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 |
|---|---|---|---|
| Add Jacobian matching to CE and activation matching | VGG-4 student distilled from VGG-9 teacher · CIFAR100, 1–100 data points/class | test accuracy (%): 12.13 -> 13.78; 26.97 -> 33.39; 33.92 -> 39.55; 46.47 -> 49.49; 50.92 -> 52.43 | Helped |
| Add Jacobian matching to CE and activation matching | VGG-4 student distilled from VGG-9 teacher · CIFAR100, 500 data points/class | test accuracy (%): 56.65 -> 54.57 | Hurt |
| Add Jacobian matching to activation and attention matching | VGG-9 transfer student with ResNet-34 teacher · MIT Scenes, all reported data sizes | test accuracy (%): 16.53 -> 18.02; 28.35 -> 29.25; 46.01 -> 47.31; 57.80 -> 58.35; 67.24 -> 67.31 | Helped |
| Add a Jacobian-norm penalty (λ = 10.0) | VGG-9 classifier · CIFAR100 with Gaussian noise, std. dev. 0.2 and 0.4 | test accuracy (%): 47.53 ± 0.23 -> 58.28 ± 0.13; 17.69 ± 0.17 -> 44.96 ± 0.19 | Helped |
| Use ℓ2 regularization instead of a Jacobian-norm penalty | VGG-9 classifier · CIFAR100 with Gaussian noise, std. dev. 0.2 and 0.4 | test accuracy (%): 58.28 ± 0.13 -> 39.93 ± 0.28; 44.96 ± 0.19 -> 13.76 ± 0.16 | Hurt |
| Match shallow rather than deeper features for Jacobian loss (ablation) | VGG-9 / ResNet-34 transfer learning · MIT Scenes, 10 data points/class | accuracy (%): 20.03 -> 22.39 | Helped |
| Use average pooling when computing Jacobians, rather than no pooling (ablation) | VGG-9 / ResNet-34 transfer learning · MIT Scenes, 10 data points/class | accuracy (%): 19.74 -> 21.87 | Helped |
| Use global average pooling instead of an s/5 window (ablation) | VGG-9 / ResNet-34 transfer learning · MIT Scenes, 10 data points/class | accuracy (%): 21.87 -> 20.00 | Hurt |
Closest papers by meaning
- Progressively Knowledge Distillation via Re-parameterizing Diffusion Reverse Process (AAAI 2024 · similarity 0.88)
- Contrastive Representation Distillation (ICLR 2020 · similarity 0.88)
- A Gift From Knowledge Distillation: Fast Optimization, Network Minimization and Transfer Learning (CVPR 2017 · similarity 0.87)
- Gap Preserving Distillation by Building Bidirectional Mappings with A Dynamic Teacher (ICLR 2025 · similarity 0.86)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.85)
- Enhancing Logits Distillation with Plug&Play Kendall's $\tau$ Ranking Loss (ICML 2025 · similarity 0.85)
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.