Information Flow in Deep Learning Classification Networks
Debanjali Banerjee, Chee‐Hung Henry Chu · 2025
Deep learning practitioners generally believe that lower layers of a CNN are less class-specific than middle layers. This study uses entropy and Jensen-Shannon Divergence to measure class-specificity at each layer of a CNN on a two-class classification task. Results confirm that class representations become more distinct in deeper layers, supporting the idea that they learn more complex and discriminative features.