Intersectional Bias in Pre-Trained Image Recognition Models

2025-06-04

AI Summary

This study investigates intersectional bias in face representations learned by ImageNet-pretrained image recognition models with respect to three sensitive attributes: age, race, and gender. We propose the first unified quantitative framework that integrates linear probe analysis, topological activation map visualization, and intersectional subgroup evaluation to systematically uncover implicit structural biases. Our analysis reveals that these models exhibit strong age discrimination and concurrently encode race and gender information—particularly among middle-aged individuals. The key contribution is the first interpretable and quantifiable assessment of three-way intersectional bias, overcoming limitations of prior work focused on single attributes or pairwise interactions. This advances fairness-aware modeling and robust representation learning for socially aware AI systems. (124 words)

Technology Category

• Computer Vision: Bias, Fairness & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

Deep Learning models have achieved remarkable success. Training them is often accelerated by building on top of pre-trained models which poses the risk of perpetuating encoded biases. Here, we investigate biases in the representations of commonly used ImageNet classifiers for facial images while considering intersections of sensitive variables age, race and gender. To assess the biases, we use linear classifier probes and visualize activations as topographic maps. We find that representations in ImageNet classifiers particularly allow differentiation between ages. Less strongly pronounced, the models appear to associate certain ethnicities and distinguish genders in middle-aged groups.

Problem

• Investigates biases in ImageNet classifiers for facial images
• Examines intersections of age, race, and gender in model biases
• Uses linear probes and topographic maps to assess bias patterns

Innovation

• Assessing biases using linear classifier probes
• Visualizing activations as topographic maps
• Analyzing intersections of age, race, gender