Saddle-To-Saddle Dynamics in Deep ReLU Networks: Low-Rank Bias in the First Saddle Escape

ICLR 2026 Conference Submission19922 Authors

Keywords: Saddle-to-Saddle, Implicit bias, Low-rank bias, Bottleneck rank

TL;DR: Starting from a small initialization, DNN escape the plateau around the origin with most of its layer being approximately rank 1.

Abstract: When a deep ReLU network is initialized with small weights, gradient descent (GD) is at first dominated by the saddle at the origin in parameter space. We study the so-called escape directions along which GD leaves the origin, which play a similar role as the eigenvectors of the Hessian for strict saddles. We show that the optimal escape direction features a \textit{low-rank bias} in its deeper layers: the first singular value of the $\ell$-th layer weight matrix is at least $\ell^{\frac{1}{4}}$ larger than any other singular value. We also prove a number of related results about these escape directions. We suggest that deep ReLU networks exhibit saddle-to-saddle dynamics, with GD visiting a sequence of saddles with increasing bottleneck rank.

Primary Area: learning theory

Submission Number: 19922