EcoSta 2024: Start Registration
View Submission - EcoSta2024
A0369
Title: Generalization and optimization of gradient methods for single-layer neural networks Authors:  Yunwen Lei - The University of Hong Kong (Hong Kong) [presenting]
Abstract: Neural networks have achieved impressive performance in various applications. The generalization and optimization of shallow neural networks (SNNs) are discussed. Both gradient descent (GD) and stochastic gradient descent (SGD) are considered for training SNNs. It shows how the generalization and optimization should be balanced to obtain consistent error bounds under a relaxed overparameterization setting. The existing estimates are improved on the weak-convexity parameter of SNNs along the trajectories of the optimization process.