Research map / Label-efficient and robust learning / Recognition and long-tail
Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers
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 C2B, a training-free, label-free pipeline that uses an LLM to propose bias attributes/captions and CLIP or Bing retrieval to build test sets for scoring biases of pre-trained visual classifiers.
Problem
Identifying biases of pre-trained visual classifiers when no labeled task-specific dataset is available, given only a textual task description.
Approach
Use an LLM to generate bias attributes, classes, and target-bias captions from the task description, retrieve images for those captions from a large database (CC12M via CLIP or Bing web search), and score biases by per-class accuracy differences of the classifier on the retrieved pseudo-labeled images.
Key results
- On CelebA, C2B (Bing) detects 12.29% of ground-truth biases vs 4.53% for B2T on FaceXFormer.
- On ImageNet-X, C2B (CC12M) detects 11.18% of ground-truth biases for ResNet50 V2 vs 2.40% for B2T.
- VQA agreement on CelebA is 0.25 for C2B (CC12M) vs 0.10 for B2T.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace B2T with C2B using Bing retrieval | FaceXFormer bias detection · CelebA | HIT (%): 4.53 -> 12.29 | Helped |
| Replace B2T with C2B using CC12M retrieval | FaceXFormer bias detection · CelebA | VQA agreement: 0.10 -> 0.25 | Helped |
| Replace B2T with C2B using CC12M retrieval | ResNet50 V2 bias detection · ImageNet-X | HIT (%): 2.40 -> 11.18 | Helped |
| Replace B2T with C2B using Bing retrieval | ViT B 16 SWAG bias detection · ImageNet-X | VQA agreement: 0.20 -> 0.32 | Helped |
| Replace B2T with C2B using Bing retrieval | FaceXFormer bias detection · CelebA | FH (%): 3.14 -> 6.88 | Hurt |
| Use CC12M retrieval instead of Bing retrieval (ablation) | C2B · ImageNet-X | HIT (%): 7.80 -> 11.18; 7.91 -> 11.21; 7.53 -> 11.10; 7.72 -> 10.85 | Helped |
| Use Bing retrieval instead of CC12M retrieval (ablation) | C2B · CelebA | HIT (%): 10.76 -> 12.29 | Helped |
| Use Bing retrieval instead of CC12M retrieval (ablation) | C2B · ImageNet-X | VQA agreement: 0.20 -> 0.26; 0.20 -> 0.28; 0.20 -> 0.28; 0.21 -> 0.32 | Helped |
Surprising findings
- On ImageNet-X, CC12M retrieval outperformed Bing in ground-truth matching, while on CelebA Bing was better.
- C2B has higher false hit rates than B2T (e.g., 6.9-7.8% vs 3.1% on CelebA), attributed to retrieval noise and broad LLM proposals.
Limitations
- Low absolute recall of ground-truth biases and higher false-hit rates than B2T.
- Retrieval noise from CLIP or web search affects the reliability of pseudo-labeled test sets.
- Evaluation relies on SBERT matching thresholds and a VQA model whose own biases may influence results.
Experimental setup
Datasets: CelebA ImageNet-1K ImageNet-X CC12M
Benchmarks: CelebA ImageNet-X
Metrics: HIT rate false hit rate miss rate VQA agreement score recall@K for retrieval
Baselines: B2T
Builds on: Bias-to-Text (B2T) CLIP Conceptual Captions 12M
Closest papers by meaning
- Discovering and Mitigating Visual Biases through Keyword Explanation (CVPR 2024 · similarity 0.89)
- Common Sense Bias Modeling for Classification Tasks (AAAI 2025 · similarity 0.89)
- Understanding Bias in Large-Scale Visual Datasets (NeurIPS 2024 · similarity 0.89)
- ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts (NeurIPS 2025 · similarity 0.88)
- Mitigating Biases in Blackbox Feature Extractors for Image Classification Tasks (NeurIPS 2024 · similarity 0.87)
- Unbiasing through Textual Descriptions: Mitigating Representation Bias in Video Benchmarks (CVPR 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.