Image Representations Learned With Unsupervised Pre-Training Contain Human-like Biases

Abstract

Recent advances in machine learning leverage massive datasets of unlabeled images from the web to learn general-purpose image representations for tasks from image classification to face recognition. But do unsupervised computer vision models automatically learn implicit patterns and embed social biases that could have harmful downstream effects? We develop a novel method for quantifying biased associations between representations of social concepts and attributes in images. We find that state-of-the-art unsupervised models trained on ImageNet, a popular benchmark image dataset curated from internet images, automatically learn racial, gender, and intersectional biases. We replicate 8 previously documented human biases from social psychology, from the innocuous, as with insects and flowers, to the potentially harmful, as with race and gender. Our results closely match three hypotheses about intersectional bias from social psychology. For the first time in unsupervised computer vision, we also quantify implicit human biases about weight, disabilities, and several ethnicities. When compared with statistical patterns in online image datasets, our findings suggest that machine learning models can automatically learn bias from the way people are stereotypically portrayed on the web.

arXiv comment10 pages, 3 figures. Replaced example image completions of real people with completions of artificial people

References in corpus (21)

• Efficient Estimation of Word Representations in Vector Space
• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• A Simple Framework for Contrastive Learning of Visual Representations
• Semantics derived automatically from language corpora contain human-like biases
• Model Cards for Model Reporting
• Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings
• Word Embeddings Quantify 100 Years of Gender and Ethnic Stereotypes
• Mitigating Bias in Algorithmic Hiring: Evaluating Claims and Practices
• Learning Representations by Maximizing Mutual Information Across Views
• Big Self-Supervised Models are Strong Semi-Supervised Learners
• The Open Images Dataset V4: Unified image classification, object detection, and visual relationship detection at scale
• Do ImageNet Classifiers Generalize to ImageNet?
• Large Scale Adversarial Representation Learning
• Bias in Bios: A Case Study of Semantic Representation Bias in a High-Stakes Setting
• Towards Fairer Datasets: Filtering and Balancing the Distribution of the People Subtree in the ImageNet Hierarchy
• Predictive Inequity in Object Detection
• Women also Snowboard: Overcoming Bias in Captioning Models
• Understanding the Origins of Bias in Word Embeddings
• Assessing Social and Intersectional Biases in Contextualized Word Representations
• Saving Face: Investigating the Ethical Concerns of Facial Recognition Auditing
• The Bottom-up Evolution of Representations in the Transformer: A Study with Machine Translation and Language Modeling Objectives

Cited by in corpus (19)

• Vector-quantized Image Modeling with Improved VQGAN
• Towards Intersectionality in Machine Learning: Including More Identities, Handling Underrepresentation, and Performing Evaluation
• Gender Bias in Word Embeddings: A Comprehensive Analysis of Frequency, Syntax, and Semantics
• WEIRD FAccTs: How Western, Educated, Industrialized, Rich, and Democratic is FAccT?
• Contrastive Language-Vision AI Models Pretrained on Web-Scraped Multimodal Data Exhibit Sexual Objectification Bias
• A Survey on Intersectional Fairness in Machine Learning: Notions, Mitigation, and Challenges
• Artificial Concepts of Artificial Intelligence: Institutional Compliance and Resistance in AI Startups
• Supervised Transfer Learning at Scale for Medical Imaging
• The Dark Side of Dataset Scaling: Evaluating Racial Classification in Multimodal Models
• Situating the social issues of image generation models in the model life cycle: a sociotechnical approach
• Nigerian Software Engineer or American Data Scientist? GitHub Profile Recruitment Bias in Large Language Models
• Mitigating Dataset Harms Requires Stewardship: Lessons from 1000 Papers
• Privacy-Preserving Kickstarting Deep Reinforcement Learning with Privacy-Aware Learners
• Social Perception of Faces in a Vision-Language Model
• Improved Transformer for High-Resolution GANs
• Privacy-preserving Object Detection
• Tractable Density Estimation on Learned Manifolds with Conformal Embedding Flows
• Interactive Discovery and Exploration of Visual Bias in Generative Text-to-Image Models
• Algorithmic Fairness Datasets: the Story so Far