Closing the Loop with Concept Regularization

Abstract

Convolutional Neural Networks (CNNs) are widely adopted in industrial settings, but are prone to biases and lack transparency. Explainable Artificial Intelligence (XAI), particularly through concept extraction (CE), allows for global explanations and bias detection, yet fails to offer corrective measures for identified biases. To bridge this gap, we introduce Concept Regularization (CoRe), which uses CE capabilities alongside human feedback to embed a regularization term during retraining. CoRe allows for the adjustments in model sensitivities based on identified biases, aligning model prediction process with expert human assessments. Our evaluations on a modified metal casting dataset demonstrate CoRe's efficacy in bias mitigation, highlighting its potential to refine models in practical applications.