Research map / Label-efficient and robust learning / Domain adaptation and generalization
Unlearning during Training: Domain-Specific Gradient Ascent for Domain Generalization
Di Zhao, Jingfeng Zhang, Hongsheng Hu, Philippe Fournier-Viger, Gillian Dobbie, Yun Sing Koh
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 |
|---|---|---|---|
| add post-epoch Identify and Unlearn (IU) | ERM · PACS | accuracy (%): 83.0 -> 85.7 | Improved |
| add EMA smoothing to the unlearning score (ablation) | ERM + IU · DomainNet | accuracy (%): 42.2 -> 43.1 | Improved |
| add IU to strong baselines | UDIM / VL2V · PACS | accuracy (%): 84.7 -> 86.6 (UDIM); 85.5 -> 87.0 (VL2V) | Improved |
| apply gradient ascent on all parameters using unlearning set (no channel selection) (ablation) | ERM · PACS | accuracy (%): 83.0 -> 78.9 | Hurt |
| apply gradient ascent on domain-specific channels over entire training set (no unlearning set selection) (ablation) | ERM · PACS | accuracy (%): 83.0 -> 76.7 | Hurt |
| combine unlearning set selection with domain-specific channel selection (ablation) | ERM · PACS | accuracy (%): 78.9 -> 85.7 | Improved |
| use Inter-Domain Variance instead of Aggregated Variance for channel selection (ablation) | ERM + IU · PACS | accuracy (%): 84.3 -> 85.7 | Improved |
| remove unlearning set from training data and retrain (instead of unlearning) (ablation) | ERM · PACS | accuracy (%): 83.0 -> 80.1 | Hurt |
| set alpha away from 1.1 (ablation) | ERM + IU · PACS | accuracy (%): 85.7 -> 83.6 (0.8); 84.1 (1.3) | Mixed |
| vary MAD threshold k from 2.0 (ablation) | ERM + IU · PACS | accuracy (%): 85.7 -> 83.4 (k=1.0); 83.3 (k=3.0) | Mixed |
| use MAD thresholding instead of fixed bottom-10% selection (ablation) | unlearning set selection · PACS | adaptivity of set size | Improved |
Closest papers by meaning
- Selective Unlearning via Representation Erasure Using Domain Adversarial Training (ICLR 2025 · similarity 0.91)
- Domain Generalization by Learning and Removing Domain-specific Features (NeurIPS 2022 · similarity 0.89)
- Towards Unsupervised Domain Generalization (CVPR 2022 · similarity 0.89)
- On Effects of Steering Latent Representation for Large Language Model Unlearning (AAAI 2025 · similarity 0.87)
- Partially Blinded Unlearning: Class Unlearning for Deep Networks from Bayesian Perspective (AAAI 2025 · similarity 0.87)
- Minimal Semantic Sufficiency Meets Unsupervised Domain Generalization (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.
