ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts

2025-10-30

AI Summary

Dataset bias is pervasive, yet existing detection methods rely on costly, fine-grained human annotations and struggle to systematically identify unknown biases. To address this, we propose ConceptScope—the first automated framework that applies sparse autoencoders to vision foundation model representations to achieve interpretable concept disentanglement. ConceptScope requires no manual labeling and decomposes visual concepts into three semantically distinct categories: target objects, contextual elements, and bias-inducing factors. It jointly leverages concept activation mapping and statistical correlation analysis for concept discovery, classification, and spatial attribution. Evaluated on Waterbirds and ImageNet, ConceptScope accurately localizes background biases and unlabeled co-occurring object biases, demonstrating strong spatial interpretability and reliable bias detection. By enabling scalable, interpretable, and zero-annotation dataset auditing, ConceptScope establishes a novel paradigm for bias-aware data curation in vision research.

Technology Category

• Computer Vision: Bias, Fairness & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Security and Privacy: Data transparency and provenance
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Dataset bias, where data points are skewed to certain concepts, is ubiquitous in machine learning datasets. Yet, systematically identifying these biases is challenging without costly, fine-grained attribute annotations. We present ConceptScope, a scalable and automated framework for analyzing visual datasets by discovering and quantifying human-interpretable concepts using Sparse Autoencoders trained on representations from vision foundation models. ConceptScope categorizes concepts into target, context, and bias types based on their semantic relevance and statistical correlation to class labels, enabling class-level dataset characterization, bias identification, and robustness evaluation through concept-based subgrouping. We validate that ConceptScope captures a wide range of visual concepts, including objects, textures, backgrounds, facial attributes, emotions, and actions, through comparisons with annotated datasets. Furthermore, we show that concept activations produce spatial attributions that align with semantically meaningful image regions. ConceptScope reliably detects known biases (e.g., background bias in Waterbirds) and uncovers previously unannotated ones (e.g, co-occurring objects in ImageNet), offering a practical tool for dataset auditing and model diagnostics.

Problem

• Automatically identifying dataset bias through visual concept analysis
• Quantifying human-interpretable concepts without manual annotations
• Detecting known and unknown biases in machine learning datasets

Innovation

• Uses Sparse Autoencoders on vision foundation models
• Categorizes concepts into target, context, and bias types
• Enables bias identification through concept-based subgrouping