This poster presents a study investigating the internal learning dynamics of deep neural networks for Alzheimer's disease classification using magnetic resonance imaging (MRI). DenseNet121 and EfficientNet-B1 models were trained on MRI datasets. Layer-wise statistical complexity, entropy, and disequilibrium were computed throughout training using the López-Ruiz–Mancini–Calbet (LMC) framework. The study examines how these information-theoretic measures evolve during learning and compares them with conventional performance metrics such as accuracy, AUC, and loss. The findings suggest that statistical complexity provides complementary insight into neural network learning dynamics and may serve as a useful tool for understanding model behavior in medical image classification.