Abstract
We study how convolutional neural networks reorganize information during learning in natural image classification tasks by tracking mutual information (MI) between inputs, intermediate representations, and labels. Across VGG-16, ResNet-18, and ResNet-50, we find that label-relevant MI grows reliably with depth while input MI depends strongly on architecture and activation, indicating that "compression'' is not a universal phenomenon. Within convolutional layers, label information becomes increasingly concentrated in a small subset of channels; inference-time knockouts, shuffles, and perturbations confirm that these high-MI channels are functionally necessary for accuracy. This behavior suggests a view of representation learning driven by selective concentration and decorrelation rather than global information reduction. Finally, we show that a simple dependence-aware regularizer based on the Hilbert-Schmidt Independence Criterion can encourage these same patterns during training, yielding small accuracy gains and consistently faster convergence.