Research map / Vision-language and multimodal / Human and embodied modeling
ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts
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.
Introduces ConceptScope, which trains sparse autoencoders on CLIP ViT-L/14 patch representations to discover visual concepts and categorize them as target, context, or bias concepts for auditing class-level dataset bias and diagnosing model robustness.
Problem
Dataset bias, where visual concepts are unevenly distributed across classes, is pervasive but hard to identify without costly fine-grained attribute annotations; existing caption-based or failure-based methods are either inconsistent or only indirectly reveal training-data bias.
Approach
Use sparse autoencoder latents from a vision foundation model as an automatically discovered, monosemantic concept dictionary; score each concept's necessity and sufficiency for a class via masked CLIP similarity to label target vs. context concepts, and flag context concepts with high class-conditional activation as bias concepts.
Key results
- SAE-based concept prediction achieves average F1 0.72 and AUPRC 0.76 across six attribute datasets, versus 0.58 F1 for LLaVA-NeXT captions and 0.43 for BLIP-2.
- On ADE20K segmentation, SAE masks reach AUPRC 0.399 vs 0.302 for LLaVA-NeXT and 0.098 for BLIP-2.
- Bias discovery Precision@10 reaches 100% on Waterbirds and CelebA and 72.9-74.0% on NICO++ variants, outperforming ViG-Bias (60.0-66.7%) on NICO++.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAE activations instead of caption-based LLaVA-NeXT attribute predictions | Concept prediction · Caltech101 | F1: 0.61{ $\pm$0.35} -> 0.83{ $\pm$0.21} | Helped |
| Use SAE activations instead of caption-based LLaVA-NeXT attribute predictions | Concept prediction · DTD | F1: 0.40{ $\pm$0.21} -> 0.57{ $\pm$0.20} | Helped |
| Use SAE activations instead of caption-based LLaVA-NeXT attribute predictions | Concept prediction · Waterbird | F1: 0.57{ $\pm$0.12} -> 0.78{ $\pm$0.07} | Helped |
| Use SAE activations instead of caption-based LLaVA-NeXT attribute predictions | Concept prediction · CelebA | F1: 0.62{ $\pm$0.24} -> 0.81{ $\pm$0.11} | Helped |
| Use SAE activations instead of caption-based LLaVA-NeXT attribute predictions | Concept prediction · RAF-DB | F1: 0.45{ $\pm$0.18} -> 0.55{ $\pm$0.18} | Helped |
| Use SAE activations instead of caption-based LLaVA-NeXT attribute predictions | Concept prediction · Stanford40 | F1: 0.80{ $\pm$0.16} -> 0.78{ $\pm$0.13} | Hurt |
| Use ConceptScope for dataset bias discovery instead of ViG-Bias | Bias discovery · Waterbirds | Precision@10: 100.0\% -> 100.0\% | No clear effect |
| Use ConceptScope for dataset bias discovery instead of ViG-Bias | Bias discovery · CelebA | Precision@10: 100.0\% -> 100.0\% | No clear effect |
| Use ConceptScope for dataset bias discovery instead of ViG-Bias | Bias discovery · NICO++ (75) | Precision@10: 60.0\% -> 72.9\% | Helped |
| Use ConceptScope for dataset bias discovery instead of ViG-Bias | Bias discovery · NICO++ (90) | Precision@10: 66.7\% -> 73.1\% | Helped |
| Use ConceptScope for dataset bias discovery instead of ViG-Bias | Bias discovery · NICO++ (95) | Precision@10: 65.0\% -> 74.0\% | Helped |
| Use SAE spatial attributions instead of LLaVA-NeXT attention maps | Concept localization · ADE20K validation | AUPRC: 0.302 -> 0.399 | Helped |
Surprising findings
- Caption-based VLM baselines perform worse than SAE concept detection largely due to linguistic variability in generic caption terms.
- Models with higher accuracy on the most frequent (typical) subgroup also tend to perform better on the rare outlier subgroup, with the same ordering on ImageNet-Sketch.
Limitations
- Discovered concepts are constrained by the knowledge encoded in CLIP representations; domain-specific datasets such as RAF-DB and DTD show lower concept prediction scores.
- Segmentation masks are patch-level (16x16) and coarse, limiting localization accuracy for small objects.
- Concept-based alignment depends on CLIP similarity, which may carry biases from pretraining; robustness was checked only via a toy experiment and CLIP variants.
Experimental setup
Datasets: ImageNet-1K Caltech-101 DTD Waterbirds CelebA RAF-DB Stanford40 ADE20K NICO++ Food101 SUN397 ImageNet-Sketch
Benchmarks: Caltech-101 DTD Waterbirds CelebA RAF-DB Stanford40 ADE20K Waterbirds bias discovery CelebA bias discovery NICO++ bias discovery ImageNet-1K ImageNet-Sketch
Metrics: F1 AUPRC precision recall Pearson correlation Spearman correlation Precision@10 silhouette score Cohen's kappa
Baselines: BLIP-2 LLaVA-NeXT DOMINO FACTS ViG-Bias
Builds on: sparse autoencoders CLIP DOMINO FACTS SSD-LLM
Closest papers by meaning
- Understanding Bias in Large-Scale Visual Datasets (NeurIPS 2024 · similarity 0.88)
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.88)
- Visual Data Diagnosis and Debiasing with Concept Graphs (NeurIPS 2024 · similarity 0.88)
- Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models (CVPR 2025 · similarity 0.85)
- ConceptBed: Evaluating Concept Learning Abilities of Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.85)
- Common Sense Bias Modeling for Classification Tasks (AAAI 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.