Feature Dynamics as Implicit Data Augmentation: A Depth-Decomposed View on Deep Neural Network Generalization
This work addresses the fundamental problem of understanding generalization in deep learning for researchers, offering a novel conceptual perspective that is incremental in linking feature dynamics to generalization.
The paper investigates why deep neural networks generalize well by analyzing the temporal consistency of internal features during training, showing that predictions remain stable when combining shallow features from earlier checkpoints with deeper ones from later stages, which acts as implicit data augmentation. It demonstrates that this consistency extends to unseen and corrupted data but collapses with random labels, and statistical tests reveal SGD injects anisotropic noise aligned with principal directions.
Why do deep networks generalize well? In contrast to classical generalization theory, we approach this fundamental question by examining not only inputs and outputs, but the evolution of internal features. Our study suggests a phenomenon of temporal consistency that predictions remain stable when shallow features from earlier checkpoints combine with deeper features from later ones. This stability is not a trivial convergence artifact. It acts as a form of implicit, structured augmentation that supports generalization. We show that temporal consistency extends to unseen and corrupted data, but collapses when semantic structure is destroyed (e.g., random labels). Statistical tests further reveal that SGD injects anisotropic noise aligned with a few principal directions, reinforcing its role as a source of structured variability. Together, these findings suggest a conceptual perspective that links feature dynamics to generalization, pointing toward future work on practical surrogates for measuring temporal feature evolution.