Discovering and Mitigating Visual Biases through Keyword Explanation

arXiv:2301.11104v4 Announce Type: replace
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 …

abstract ai deployments arxiv bias biases computer computer vision cs.cv cs.lg deployments however human interpretation issue nature real-world ai sample statistics supervision text through type vision vision models visual visualization world