[Paper Review] Discovering and Mitigating Visual Biases through Keyword Explanation

Younghyun Kim, Sangwoo Mo+4 more|arXiv (Cornell University)|Jan 26, 2023

Multimodal Machine Learning Applications4 citations

TL;DR

This paper introduces the Bias-to-Text (B2T) framework, which uses language interpretation from image captions to discover and mitigate visual biases in vision models. By analyzing keywords in captions of mispredicted or generated images and applying CLIP and SD score functions, B2T identifies spurious correlations—such as 'player' with 'female' or 'stethoscope' with 'nurse'—and improves worst-group accuracy by 11% on Waterbirds and reduces unfair/unsafe generations in text-to-image models.

ABSTRACT

Addressing biases in computer vision models is crucial for real-world AI deployments. However, mitigating visual biases is challenging due to their unexplainable nature, often identified indirectly through visualization or sample statistics, which necessitates additional human supervision for interpretation. To tackle this issue, we propose the Bias-to-Text (B2T) framework, which interprets visual biases as keywords. Specifically, we extract common keywords from the captions of mispredicted images to identify potential biases in the model. We then validate these keywords by measuring their similarity to the mispredicted images using a vision-language scoring model. The keyword explanation form of visual bias offers several advantages, such as a clear group naming for bias discovery and a natural extension for debiasing using these group names. Our experiments demonstrate that B2T can identify known biases, such as gender bias in CelebA, background bias in Waterbirds, and distribution shifts in ImageNet-R/C. Additionally, B2T uncovers novel biases in larger datasets, such as Dollar Street and ImageNet. For example, we discovered a contextual bias between "bee" and "flower" in ImageNet. We also highlight various applications of B2T keywords, including debiased training, CLIP prompting, and model comparison.

Motivation & Objective

• To address the challenge of diagnosing and mitigating unknown visual biases in vision models without relying on pre-defined bias categories.
• To enable explainable bias discovery by generating natural language descriptions from images and extracting common keywords.
• To develop a universal, model-agnostic framework applicable to both image classifiers and text-to-image generative models.
• To improve worst-group robustness in zero-shot classifiers and reduce unfair or unsafe image generation in diffusion models.
• To provide a practical, scalable method for debiasing models using language-based interpretation of visual bias patterns.

Proposed method

• Use a pre-trained captioning model (e.g., CLIP-Cap) to generate natural language descriptions from potentially biased images, such as mispredicted images or text-conditioned generated images.
• Extract common keywords from these captions to identify recurring terms associated with biased predictions or generations.
• Propose a CLIP score to measure similarity between bias keywords and mispredicted images, ensuring alignment with model failures.
• Introduce an SD (Stable Diffusion) score to verify if generated images reflect bias keywords by comparing diffusion scores of original prompts and bias keywords.
• Use the identified keywords to debias models: augment prompts for zero-shot classifiers or guide diffusion scores in text-to-image models.
• Apply the framework to detect and mitigate spurious correlations in both image classification and text-to-image generation tasks.

Experimental results

Research questions

• RQ1 Can language interpretation of image captions effectively uncover unknown visual biases in vision models?
• RQ2 How can score functions like CLIP and SD be used to filter and validate bias keywords from captioned images?
• RQ3 To what extent can keyword-based debiasing improve worst-group accuracy in zero-shot image classifiers?
• RQ4 Can the B2T framework detect and mitigate unfair or unsafe attributes in text-to-image generation, such as gender or nudity bias?
• RQ5 Is the B2T framework generalizable across different vision models, including both discriminative and generative architectures?

Key findings

• The B2T framework discovered a new spurious correlation between 'player' and 'female' in the Kaggle Face dataset, which was previously undetected.
• Debiasing using B2T improved worst-group accuracy on the Waterbirds dataset by 11% compared to the baseline.
• For text-to-image models, B2T successfully detected gender biases—such as 'nurse' being linked to 'woman' or 'stethoscope'—and unsafe biases like 'naked' in generated images.
• The SD score effectively identified unsafe or unfair attributes in generated images, with lower scores indicating stronger bias presence.
• The framework demonstrated robust performance across diverse benchmarks, including ImageNet-C and various text-to-image prompts, showing consistent bias detection.
• The use of CLIP and SD scores significantly reduced false positive bias keywords, improving the reliability of discovered biases.