Bias Mitigation in Visual Question Answering: Implementation and Analysis

This study explores the influence of social biases—such as gender, racial, and geographical biases—on Visual Question Answering (VQA) models, with a focus on how these biases affect image-based reasoning. By systemati- cally comparing standard VQA models with those incor- porating bias-mitigation strategies, we assess the effec- tiveness of these techniques in promoting fairness and re- ducing biased outputs. Leveraging a combination of bias- sensitive datasets and diverse image sources, this research provides a comprehensive analysis of model behavior un- der varying bias conditions. The ultimate goal is to iden- tify robust methods for mitigating bias in VQA systems, thereby advancing the development of more equitable and trustworthy AI models