Identifying Implicit Social Biases in Vision-Language Models

TL;DR

This paper systematically analyzes social biases in CLIP, revealing harmful associations between demographic groups and stereotypes, and emphasizes the need for bias mitigation and dataset transparency in vision-language models.

Contribution

It introduces a comprehensive taxonomy of social biases in vision-language models and demonstrates how these biases manifest in CLIP's retrieval results, highlighting sources and implications.

Findings

CLIP often retrieves images with stereotypical associations.

Harmful biases are present in the training datasets.

Biases can be traced back to the source datasets.

Abstract

Vision-language models, like CLIP (Contrastive Language Image Pretraining), are becoming increasingly popular for a wide range of multimodal retrieval tasks. However, prior work has shown that large language and deep vision models can learn historical biases contained in their training sets, leading to perpetuation of stereotypes and potential downstream harm. In this work, we conduct a systematic analysis of the social biases that are present in CLIP, with a focus on the interaction between image and text modalities. We first propose a taxonomy of social biases called So-B-IT, which contains 374 words categorized across ten types of bias. Each type can lead to societal harm if associated with a particular demographic group. Using this taxonomy, we examine images retrieved by CLIP from a facial image dataset using each word as part of a prompt. We find that CLIP frequently displays undesirable associations between harmful words and specific demographic groups, such as retrieving mostly pictures of Middle Eastern men when asked to retrieve images of a "terrorist". Finally, we conduct an analysis of the source of such biases, by showing that the same harmful stereotypes are also present in a large image-text dataset used to train CLIP models for examples of biases that we find. Our findings highlight the importance of evaluating and addressing bias in vision-language models, and suggest the need for transparency and fairness-aware curation of large pre-training datasets.