Learning Dynamics of Deep Linear Networks Beyond the Edge of Stability
This work provides theoretical insights into the edge of stability phenomenon in deep learning, explaining why shallow models or simple tasks may not exhibit it and how oscillations are confined to top features, which is incremental for understanding training dynamics.
The paper analyzes the learning dynamics of deep linear networks beyond the edge of stability, showing that loss oscillations follow a period-doubling route to chaos and occur within a small subspace characterized by the learning rate, with experiments supporting the theory.
Deep neural networks trained using gradient descent with a fixed learning rate $η$ often operate in the regime of "edge of stability" (EOS), where the largest eigenvalue of the Hessian equilibrates about the stability threshold $2/η$. In this work, we present a fine-grained analysis of the learning dynamics of (deep) linear networks (DLNs) within the deep matrix factorization loss beyond EOS. For DLNs, loss oscillations beyond EOS follow a period-doubling route to chaos. We theoretically analyze the regime of the 2-period orbit and show that the loss oscillations occur within a small subspace, with the dimension of the subspace precisely characterized by the learning rate. The crux of our analysis lies in showing that the symmetry-induced conservation law for gradient flow, defined as the balancing gap among the singular values across layers, breaks at EOS and decays monotonically to zero. Overall, our results contribute to explaining two key phenomena in deep networks: (i) shallow models and simple tasks do not always exhibit EOS; and (ii) oscillations occur within top features. We present experiments to support our theory, along with examples demonstrating how these phenomena occur in nonlinear networks and how they differ from those which have benign landscape such as in DLNs.