Research map / Label-efficient and robust learning / Recognition and long-tail
Training Debiased Subnetworks With Contrastive Weight Pruning
Geon Yeong Park, Sangmin Lee, Sang Wan Lee, Jong Chul Ye
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 DCWP instead of the strongest baseline | debiased classifiers · CMNIST, bias ratio 0.5% | unbiased accuracy: 86.74% → 93.41% | Improved |
| use DCWP instead of the strongest baseline | debiased classifiers · CIFAR10-C, bias ratio 0.5% | unbiased accuracy: 27.86% → 35.90% | Improved |
| add pruning using trained masks (ablation) | DCWP · CMNIST, bias ratio 1% | accuracy: +7.19% | Improved |
| add pruning using trained masks (ablation) | DCWP with weighted cross-entropy · CMNIST, bias ratio 1% | accuracy: +11.59% | Improved |
| add pruning using trained masks (ablation) | DCWP with alignment loss and weighted cross-entropy · CMNIST, bias ratio 1% | accuracy: +8.68% | Improved |
| add alignment loss (ablation) | pruned DCWP trained with weighted cross-entropy · CMNIST, bias ratio 1% | accuracy | Improved |
| add weighted cross-entropy for bias-conflicting samples (ablation) | unpruned pretrained weights · CMNIST, bias ratio 1% | accuracy | Improved |
| add alignment loss (ablation) | unpruned pretrained weights trained with weighted cross-entropy · CMNIST, bias ratio 1% | accuracy | Improved |
| remove pruning (ablation) | DCWP · CelebA (Blonde) | worst-group accuracy | Hurt |
| remove pruning (ablation) | DCWP · CelebA (Blonde) | average accuracy | No clear effect |
| use early-stopped ERM to mine bias-conflicting samples instead of the dataset-tailored mining model (ablation) | DCWP · CIFAR10-C, bias ratio 1% | unbiased accuracy | Mixed |
| fine-tune surviving weights without resetting them to initialization (ablation) | DCWP · CMNIST, bias ratio 1% | generalization performance | Improved |
Closest papers by meaning
- Bias In, Bias Out? Finding Unbiased Subnetworks in Vanilla Models (CVPR 2026 · similarity 0.92)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.90)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.86)
- Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair (CVPR 2024 · similarity 0.85)
- Initial Guessing Bias: How Untrained Networks Favor Some Classes (ICML 2024 · similarity 0.85)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 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.
