Research map / Label-efficient and robust learning / Recognition and long-tail
Rank-Guided Pseudo-Bias Learning for Robust Black-Box Adaptation
Rajeev Ranjan Dwivedi, Anshuman Dangwal, Vinod K Kurmi
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 supervised contrastive loss with pseudo-bias labels to CE (ablation) | adapter on frozen ResNet-18 · Waterbirds | worst-group accuracy: 38.90 -> 66.67 | Improved |
| add contrastive loss to CE (ablation) | adapter on frozen ResNet-18 · CMNIST-0.9 | bias-conflicting accuracy: 61.72 -> 84.56 | Improved |
| add contrastive loss to CE (ablation) | adapter on frozen ResNet-18 · CelebA | worst-group accuracy: 27.20 -> 64.44 | Improved |
| add contrastive loss to CE (ablation) | adapter on frozen ResNet-18 · CMNIST-0.995 | bias-conflicting accuracy: 49.09 -> 52.73 | Improved |
| add contrastive loss on top of cluster margin (CM) (ablation) | CM baseline · Waterbirds | worst-group accuracy: 80.29 -> 83.01 | Improved |
| add contrastive loss on top of cluster margin (CM) (ablation) | CM baseline · CelebA | worst-group accuracy: 80.79 -> 83.10 | Improved |
| use oracle true bias labels instead of pseudo labels (ablation) | PLD-Debias · Waterbirds | worst-group accuracy: 83.02 -> 85.81 | Improved |
| use oracle true bias labels instead of pseudo labels (ablation) | PLD-Debias · CelebA | worst-group accuracy: 83.11 -> 83.33 | No clear effect |
| set contrastive weight to 0.01 rather than 0 (ablation) | PLD-Debias · Waterbirds (ResNet-18) | worst-group accuracy: 80.85 -> 83.02 | Improved |
| increase contrastive weight to 1 or more (ablation) | PLD-Debias · Waterbirds (ResNet-18) | worst-group accuracy: 83.02 -> 66.04 (lambda 0.01 -> 1); 52.59 at lambda 10 | Hurt |
| lower spurious correlation strength (bias ratio 0.8 -> 0.5) with pseudo labels | PLD-Debias · CMNIST ResNet-18 | pseudo-bias label accuracy: 89.64 -> 60.14 | Hurt |
| use PLD-Debias at weak bias ratio 0.5 versus CM | PLD-Debias · CMNIST-0.5 ResNet-18 | bias-conflicting accuracy: 84.54 -> 87.16 | Improved |
Closest papers by meaning
- Mitigating Biases in Blackbox Feature Extractors for Image Classification Tasks (NeurIPS 2024 · similarity 0.86)
- GS-Bias: Global-Spatial Bias Learner for Single-Image Test-Time Adaptation of Vision-Language Models (ICML 2025 · similarity 0.85)
- A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks (NeurIPS 2024 · similarity 0.85)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.85)
- Interpretable Debiasing of Vision-Language Models for Social Fairness (CVPR 2026 · similarity 0.85)
- BendVLM: Test-Time Debiasing of Vision-Language Embeddings (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.
