Research map / Label-efficient and robust learning / Knowledge distillation
Forget Me Not: Fighting Local Overfitting With Knowledge Fusion and Distillation
Uri Stern, Eli Corn, Daphna Weinshall
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 |
|---|---|---|---|
| apply Knowledge Fusion of mid-training checkpoints | ResNet18 · CIFAR-100 (clean) | accuracy: 78.85 -> 79.36 | Improved |
| apply Knowledge Fusion of mid-training checkpoints | ResNet18 · TinyImageNet (clean) | accuracy: 65.30 -> 68.33 | Improved |
| distill the KF ensemble into a single student (ablation) | KF ensemble · CIFAR-100 (clean) | accuracy: 79.36 -> 80.29 | Improved |
| distill the KF ensemble into a single student (ablation) | KF ensemble · TinyImageNet (clean) | accuracy: 68.33 -> 69.96 | Improved |
| distill into a student instead of averaging checkpoint weights (ablation) | KF · CIFAR-100 (clean) | accuracy: 78.48 -> 80.29 | Improved |
| apply KF under heavy label noise | ResNet18 · CIFAR-100 (40% asym noise) | accuracy: 49.95 -> 63.37 | Improved |
| distill the KF ensemble (ablation) | KF ensemble · CIFAR-100N (40% noise) | accuracy: 63 -> 64.12 | Improved |
| distill the independent ensemble (ablation) | independent ensemble · CIFAR-100 (clean) | accuracy: 82.13 -> 79.93 | Hurt |
| distill the independent ensemble (ablation) | independent ensemble · CIFAR-100 (40% asym noise) | accuracy: 55.18 -> 56.84 | Improved |
| replace horizontal ensemble with KF | ResNet18 · CIFAR-100 (clean) | accuracy: 78.42 -> 79.36 | Improved |
| apply KF to a ViT-B/16 | ViT-B/16 · ImageNet | accuracy: 79.16 -> 80.34 | Improved |
| apply KF on head-only finetuning of pretrained model | ResNet18 · CIFAR-100 (transfer) | accuracy: 61.7 -> 65.24 | Improved |
Closest papers by meaning
- On Local Overfitting and Forgetting in Deep Neural Networks (AAAI 2025 · similarity 0.97)
- On the Over-Memorization During Natural, Robust and Catastrophic Overfitting (ICLR 2024 · similarity 0.88)
- The Effectiveness of Random Forgetting for Robust Generalization (ICLR 2024 · similarity 0.87)
- Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning (NeurIPS 2026 · similarity 0.86)
- Label-Agnostic Forgetting: A Supervision-Free Unlearning in Deep Models (ICLR 2024 · similarity 0.86)
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples (NeurIPS 2025 · 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.
