Thresholding Procedure via Barzilai-Borwein Rules for the Steplength Selection in Stochastic Gradient Methods

A crucial aspect in designing a learning algorithm is the selection of the hyperparameters (parameters that are not trained during the learning process). In particular the effectiveness of the stochastic gradient methods strongly depends on the steplength selection. In recent papers [9, 10], Franchini et al. propose to adopt an adaptive selection rule borrowed from the full-gradient scheme known as Limited Memory Steepest Descent method [8] and appropriately tailored to the stochastic framework. This strategy is based on the computation of the eigenvalues (Ritz-like values) of a suitable matrix obtained from the gradients of the most recent iterations, and it enables to give an estimation of the local Lipschitz constant of the current gradient of the objective function, without introducing line-search techniques. The possible increase of the size of the sub-sample used to compute the stochastic gradient is driven by means of an augmented inner product test approach [3]. The whole procedure makes the tuning of the parameters less expensive than the selection of a fixed steplength, although it remains dependent on the choice of threshold values bounding the variability of the steplength sequences. The contribution of this paper is to exploit a stochastic version of the Barzilai-Borwein formulas [1] to adaptively select the endpoints range for the Ritz-like values. A numerical experimentation for some convex loss functions highlights that the proposed procedure remains stable as well as the tuning of the hyperparameters appears less expensive