Bi-Source Class Visualization: An Adversarial Neural Network-based Approach for Unbiased Class Visualization

This paper introduces Bi-Source Class Visualization (BSCV), a novel approach for generating interpretable and visually coherent class visualizations using adversarial neural networks. Unlike existing methods that rely on hand-crafted regularization terms, BSCV incorporates two sources of teaching signals: the loss derived from the activation maximization cost function of a specific logit in a trained classifier, and the adversarial loss obtained from a discriminator network trained on the same dataset. By optimizing the visualizations based on both the classifier and discriminator, BSCV achieves robust and less biased class visualizations. Experimental evaluation demonstrates the efficacy of BSCV in generating visually appealing and discriminative visualizations. Comparison with existing methods highlights the advantages of BSCV in capturing the body of the subject and producing higher contrast visualizations. The paper also discusses future directions for evaluation metrics and explores the limitations and potential applications of BSCV. Overall, BSCV offers a promising approach for understanding and interpreting the internal representations of deep neural networks.