Stochastic Recursive Gradient Algorithm for Nonconvex Optimization
This work addresses optimization challenges in machine learning for nonconvex losses, but it is incremental as it builds on existing SARAH methods.
The paper tackles the problem of nonconvex optimization by analyzing a mini-batch version of the SARAH algorithm, achieving a sublinear convergence rate for general nonconvex functions and a linear convergence rate for gradient dominated functions, with advantages over other stochastic gradient methods.
In this paper, we study and analyze the mini-batch version of StochAstic Recursive grAdient algoritHm (SARAH), a method employing the stochastic recursive gradient, for solving empirical loss minimization for the case of nonconvex losses. We provide a sublinear convergence rate (to stationary points) for general nonconvex functions and a linear convergence rate for gradient dominated functions, both of which have some advantages compared to other modern stochastic gradient algorithms for nonconvex losses.