The authors investigate the behavior of gradient descent in high-dimensional landscapes, revealing a transition from informative to uninformative local curvature during optimization. Successful recovery is achieved before the algorithmic transition at large dimensions.
Modified gradient descent methods achieve global L2 cost minimization in overparametrized and underparametrized settings.
新しい局所最適性の境界を開発する勾配降下法(GD)に関する研究。
This research paper investigates the properties and optimization behavior of strongly quasiconvex functions, demonstrating the exponential convergence of first and second-order gradient methods in minimizing these functions without relying on the typical Lipschitz continuity assumption.