Research map / Label-efficient and robust learning / OOD and anomaly detection
General Greedy De-Bias Learning
Xinzhe Han, Shuhui Wang, Chi Su, Qingming Huang, Qi Chuan Tian
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 curriculum regularization (GGD_cr) instead of constant greedy removal (GGD_gs) (ablation) | GGD dq · VQA-CP v2 test | accuracy: 57.12 -> 59.57 | Improved |
| add distribution bias to shortcut bias (ablation) | GGD_cr · VQA-CP v2 test | accuracy: 55.81 -> 59.57 | Improved |
| add shortcut bias to distribution bias (ablation) | GGD_gs · VQA-CP v2 test | accuracy: 48.27 -> 57.12 | Improved |
| use greedy learning instead of summing biased and base outputs (ablation) | UpDn · VQA-CP v2 test | accuracy: 35.46 -> 57.12 | Improved |
| use greedy GGD_cr instead of LMH+RUBi combination (ablation) | UpDn · VQA-CP v2 test | accuracy: 51.54 -> 59.57 | Improved |
| relax lambda_t from 1.0 to 0.95 (ablation) | GGD_gs · Biased-MNIST (rho_train 0.999) | accuracy: in-distribution 91.42 -> 99.96; OOD 68.31 -> 58.79 | Mixed |
| anneal lambda_t (curriculum) instead of constant 1 (ablation) | GGD_gs · Biased-MNIST (rho_train 0.999) | accuracy: in-distribution 91.42 -> 99.58; OOD 68.31 -> 67.01 | Improved |
| add self-ensemble GGD_cr without predefined bias (ablation) | UpDn · VQA-CP v2 test | accuracy: 39.89 -> 54.42 | Improved |
| remove distribution bias before self-ensemble (ablation) | GGD_cr^se · VQA-CP v2 test | accuracy: 54.42 -> 57.08 | Improved |
| use GGD_cr^1k instead of ReBias^1k (ablation) | ResNet-18 · Biased-MNIST (rho_train 0.999, rho_test 0) | accuracy: 32.84 -> 68.36 | Improved |
| use constant greedy GGD_gs dq on in-distribution data | UpDn · VQA v2 val | accuracy: 63.79 -> 59.30 | Hurt |
| add GGD_cr with extra AddSentOne hard examples | BiDAF · Adversarial SQuAD AddSent | F1: 50.11 -> 53.94 | Improved |
Closest papers by meaning
- Can Biases in ImageNet Models Explain Generalization? (CVPR 2024 · similarity 0.86)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.86)
- A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample Perspective (NeurIPS 2024 · similarity 0.85)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.85)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.85)
- Changing the Training Data Distribution to Reduce Simplicity Bias Improves In-distribution Generalization (NeurIPS 2024 · 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.
