부산대학교 도서관

The generalization capacity of deep neural networks has been studied in a variety of ways, including at least two distinct categories of approaches: one based on the shape of the loss landscape in parameter space, and the other based on the structure of the representation manifold in feature space (that is, in the space of unit activities). Although these two approaches are related, they are rarely studied together explicitly. Here, we present an analysis that bridges this gap. We show that in the final phase of learning in deep neural networks, the compression of the manifold of neural representations correlates with the flatness of the loss around the minima explored by SGD. This correlation is predicted by a relatively simple mathematical relationship: a flatter loss corresponds to a lower upper bound on the compression metrics of neural representations. Our work builds upon the linear stability insight by Ma and Ying, deriving inequalities between various compression metrics and quantities involving sharpness. Empirically, our derived inequality predicts a consistently positive correlation between representation compression and loss sharpness in multiple experimental settings. Overall, we advance a dual perspective on generalization in neural networks in both parameter and feature space.