Nonparametric Budgeted Stochastic Gradient Descent

Abstract

One of the most challenging problems in kernel online learning is to bound the model size. Budgeted kernel online learning addresses this issue by bounding the model size to a predefined budget. However, determining an appropriate value for such predefined budget is arduous. In this paper, we propose the Nonparametric Budgeted Stochastic Gradient Descent that allows the model size to automatically grow with data in a principled way. We provide theoretical analysis to show that our framework is guaranteed to converge for a large collection of loss functions (e.g. Hinge, Logistic, L2, L1, and \varepsilon-insensitive) which enables the proposed algorithm to perform both classification and regression tasks without hurting the ideal convergence rate O\left(\frac1T\right) of the standard Stochastic Gradient Descent. We validate our algorithm on the real-world datasets to consolidate the theoretical claims.