Research map / Vision-language and multimodal / Human and embodied modeling
Discovering and Mitigating Visual Biases through Keyword Explanation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Bias-to-Text (B2T), which explains visual classifier biases as keywords by captioning mispredicted images, extracting common keywords, and validating them with CLIP-based scores, then uses them for debiasing, prompting, and model comparison.
Problem
Visual biases in image classifiers are hard to discover and name; prior methods rely on visualizations or sample statistics needing human interpretation, or are limited to pre-defined bias vocabularies.
Approach
Interpret visual biases as keywords by aggregating common keywords from captions of mispredicted images, then validate each keyword with a CLIP score that contrasts similarity to mispredicted vs. correctly predicted images; the keyword form enables group naming and DRO-based debiasing.
Key results
- DRO-B2T reaches 90.4% worst-group accuracy on CelebA blond and 90.7% on Waterbirds, outperforming unsupervised debiasing baselines such as JTT (81.5%, 83.8%) and CNC (88.8%, 88.5%).
- CLIP zero-shot with B2T-pos keywords raises Waterbirds worst-group accuracy from 50.3% to 61.7% and CelebA blond from 76.2% to 80.0%.
- CLIP score correlates with subgroup AUROC at -0.95 on the Waterbirds waterbird class, indicating higher scores mark stronger biases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Infer bias labels from B2T keywords and use them for DRO training | DRO-B2T · CelebA blond | worst-group accuracy: 88.8±0.9 -> 90.4±0.9 | Helped |
| Infer bias labels from B2T keywords and use them for DRO training | DRO-B2T · Waterbirds | worst-group accuracy: 88.5±0.3 -> 90.7±0.3 | Helped |
| Add B2T keywords with positive CLIP scores to prompts | CLIP zero-shot · CelebA blond | worst-group accuracy: 76.2 -> 80.0 | Helped |
| Add B2T keywords with positive CLIP scores to prompts | CLIP zero-shot · Waterbirds | worst-group accuracy: 50.3 -> 61.7 | Helped |
| Add B2T keywords with negative CLIP scores to prompts (ablation) | CLIP zero-shot · Waterbirds | worst-group accuracy: 50.3 -> 45.4 | Hurt |
| Use different captioning models to generate B2T descriptions (ablation) | B2T · CelebA blond and Waterbirds | bias keyword agreement | No clear effect |
| Use different vision-language scoring models for CLIP scores (ablation) | B2T · CelebA blond and Waterbirds | keyword score ranking | No clear effect |
| Use B2T keywords with positive rather than negative CLIP scores in prompts (ablation) | CLIP zero-shot · CelebA blond | worst-group accuracy: 72.9 -> 80.0 | Mixed |
Surprising findings
- B2T-neg keywords (non-bias keywords) reduced CLIP zero-shot worst-group accuracy on Waterbirds (50.3% to 45.4%), showing keyword choice matters for prompting.
- DRO-B2T with inferred bias labels surpasses DRO trained with ground-truth bias labels, possibly due to noise in GT annotations.
Limitations
- Relies on pre-trained captioning and scoring models that may fail in uncommon domains such as medical or satellite images.
- Scoring models may not capture image-text similarity well due to training data limitations.
- Discovered keywords are the opposite of the biased concept in some cases and require human interpretation for final judgment.
Experimental setup
Datasets: CelebA Waterbirds ImageNet ImageNet-R ImageNet-C Dollar Street
Benchmarks: CelebA Waterbirds ImageNet ImageNet-R ImageNet-C Dollar Street
Metrics: worst-group accuracy average accuracy AUROC of subgroup accuracy CLIP score
Baselines: ERM LfF GEORGE JTT CNC Domino Failure Direction DRO CLIP zero-shot prompting with group names
Builds on: Distributionally robust optimization (group DRO) ClipCap YAKE CLIP
Closest papers by meaning
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.89)
- Common Sense Bias Modeling for Classification Tasks (AAAI 2025 · similarity 0.88)
- Understanding Bias in Large-Scale Visual Datasets (NeurIPS 2024 · similarity 0.88)
- Implicit Bias Injection Attacks against Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.87)
- Joint Vision-Language Social Bias Removal for CLIP (CVPR 2025 · similarity 0.87)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.87)
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.