Gradient Descent Only Converges to Minimizers: Non-Isolated Critical Points and Invariant Regions
read the original abstract
Given a non-convex twice differentiable cost function f, we prove that the set of initial conditions so that gradient descent converges to saddle points where \nabla^2 f has at least one strictly negative eigenvalue has (Lebesgue) measure zero, even for cost functions f with non-isolated critical points, answering an open question in [Lee, Simchowitz, Jordan, Recht, COLT2016]. Moreover, this result extends to forward-invariant convex subspaces, allowing for weak (non-globally Lipschitz) smoothness assumptions. Finally, we produce an upper bound on the allowable step-size.
This paper has not been read by Pith yet.
Forward citations
Cited by 2 Pith papers
-
Limitations of Lazy Training of Two-layers Neural Networks
For quadratic targets in d dimensions, two-layer quadratic networks achieve lower risk when fully trained than in random features or neural tangent regimes if hidden units < d.
-
Accelerated Gradient Methods for Nonconvex Optimization: Escape Trajectories From Strict Saddle Points and Convergence to Local Minima
Theoretical analysis of accelerated gradient methods showing almost-sure escape from strict saddles and linear exit times, plus a subclass achieving near-optimal convergence to local minima in convex neighborhoods of ...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.