A Prompt Array Keeps the Bias Away: Debiasing Vision-Language Models with Adversarial Learning

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 806–822 November 20–23, 2022. ©2022 Association for Computational Linguistics 806 A Prompt Array Keeps the Bias Away: Debiasing Vision-Language Models with Adversarial Learning Hugo Berg∗, Siobhan Mackenzie Hall, Yash Bhalgat, Hannah Rose Kirk, Aleksandar Shtedritski, Max Bain Oxford Artificial Intelligence Society, University of Oxford Abstract Vision-language models can encode societal biases and stereotypes, but there are challenges to measuring and mitigating these multimodal harms due to lacking measurement robustness and feature degradation. To address these chal- lenges, we investigate bias measures and apply ranking metrics for image-text representations. We then investigate debiasing methods and show that prepending learned embeddings to text queries that are jointly trained with adver- sarial debiasing and a contrastive loss reduces various bias measures with minimal degrada- tion to the image-text representation. 1 Introduction Large-scale, pretrained vision-language (VL) mod- els are growing in popularity due to their impres- sive performance on downstream tasks with mini- mal finetuning. Their success can be attributed to three main advances: the rise of transformers in natural language processing (NLP) (Devlin et al., 2018), cross-modal contrastive learning (Zhai and Wu, 2018) and the availability of large multimodal web datasets (Changpinyo et al., 2021). These models, including CLIP (Radford et al., 2021), are readily available through APIs (Evertrove; Hug- gingFace), allowing non-technical users to capital- ize on their high performance ‘out of the box’ on zero-shot tasks (Kirk et al., 2021). Despite these benefits, an expansion in scope for downstream applications comes with greater risk of perpetuating damaging biases that the models learn during pretraining on web-scraped datasets which are too large to be manually audited for quality (Birhane et al., 2021). Cultural and temporal speci- ficity is also of concern given models are trained on a snapshot in space and time (Haraway, 2004), thus reinforcing negative stereotypes that may otherwise naturally alter through societal pressures and norm change. ∗Corresponding author: hugo@hugob.se The risk and type of societal harm intimately in- teracts with the downstream task at hand. Clearly, using VL models for dog-species classification poses very different dangers to projecting the sim- ilarity of human faces onto axes of criminality (Wu and Zhang, 2016; Fussell, 2020) or homo- sexuality (Wang and Kosinski, 2018). Applications of this kind are extremely hard to ethically moti- vate and there may be no appropriate use case that justifies their associated risks. Even in more be- nign applications such as image search, there may be harmful consequences arising from representa- tional and/or allocational harms. Representational harms come from the technological entrenchment of stereotypical perceptions; for instance, the over- representation of one gender when querying for a profession (e.g., “nurse” versus “doctor”) or one ethnicity in explicit and NSFW content (Birhane et al., 2021). Allocational harms arise when an individual’s or group’s access to resources and op- portunity are differentially impacted (Weidinger et al., 2021); for instance, …

Content truncated to 500 words