Research map / Label-efficient and robust learning / Recognition and long-tail
Gradient Extrapolation for Debiased Representation Learning
Ihab Asaad, Maha Shadaydeh, Joachim Denzler
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 |
|---|---|---|---|
| raise extrapolation factor beta from -1 (equivalent to biased loss) to 0 and 1 (ablation) | GERNE · C-MNIST 0.5% minority (known) | unbiased test accuracy | Improved |
| use beta=1 instead of beta=0 (ablation) | GERNE · C-MNIST 0.5% minority (known) | generalization | Improved |
| increase beta to 1.2, near upper bound 1.22 (ablation) | GERNE · C-MNIST 0.5% minority (known) | training stability | Hurt |
| use GERNE extrapolated gradient instead of resampling | Resampling · bFFHQ (known attributes) | accuracy: 72.13 -> 85.20 | Improved |
| use GERNE instead of Group DRO | Group DRO · C-MNIST (1% minority, known) | accuracy: 68.78 -> 84.47 | Improved |
| use GERNE instead of Group DRO | Group DRO · C-CIFAR-10 (5% minority, known) | accuracy: 57.32 -> 62.40 | Improved |
| use GERNE instead of LC with unknown attributes | LC · C-MNIST (5% minority, unknown) | accuracy: 91.16 -> 90.98 | Hurt |
| lower pseudo-minority threshold t to 5e-4 (about 28 samples) (ablation) | GERNE unknown attrs · C-MNIST 0.5% minority | test accuracy: approximately 70% | Mixed |
| use higher threshold t=1e-2 for pseudo-minority group (ablation) | GERNE unknown attrs · C-MNIST 0.5% minority | best beta | Hurt |
| use group-balanced validation set versus validation without attributes (ablation) | GERNE · CelebA | worst-group accuracy: 86.28 -> 74.24 | Hurt |
| train GERNE on validation set instead of DFR last-layer training | DFR · CelebA | worst-group test accuracy: 86.30 -> 90.97 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DeNetDM — DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.86)
- SEBRA : Debiasing through Self-Guided Bias Ranking (ICLR 2025 · similarity 0.85)
- Twofold Debiasing Enhances Fine-Grained Learning with Coarse Labels (AAAI 2025 · similarity 0.84)
- Limits to scalable evaluation at the frontier: LLM as judge won’t beat twice the data (ICLR 2025 · similarity 0.84)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.84)
- Improving Subgroup Robustness via Data Selection (NeurIPS 2024 · similarity 0.83)
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.
