On the Regularizing Property of Stochastic Gradient Descent

Abstract

Stochastic gradient descent is one of the most successful approaches for solving large-scale problems, especially in machine learning and statistics. At each iteration, it employs an unbiased estimator of the full gradient computed from one single randomly selected data point. Hence, it scales well with problem size and is very attractive for truly massive dataset, and holds significant potentials for solving large-scale inverse problems. In the recent literature of machine learning, it was empirically observed that when equipped with early stopping, it has regularizing property. In this work, we rigorously establish its regularizing property (under \textit{a priori} early stopping rule), and also prove convergence rates under the canonical sourcewise condition, for minimizing the quadratic functional for linear inverse problems. This is achieved by combining tools from classical regularization theory and stochastic analysis. Further, we analyze the preasymptotic weak and strong convergence behavior of the algorithm. The theoretical findings shed insights into the performance of the algorithm, and are complemented with illustrative numerical experiments.

22 pages, better presentation

Cited by in corpus (4)

• Stochastic asymptotical regularization for linear inverse problems
• Stochastic mirror descent method for linear ill-posed problems in Banach spaces
• An Analysis of Stochastic Variance Reduced Gradient for Linear Inverse Problems
• Randomized progressive iterative approximation for B-spline curve and surface fittings