PROJECT TITLE :

Scaling Up Generalized Kernel Methods

ABSTRACT:

Over the course of the past two decades, kernel methods have enjoyed a great deal of success. During this era of big data, the collection of data has experienced a tremendous expansion. The currently available kernel methods, on the other hand, do not scale well enough, both during the training and the predicting steps. In order to find a solution to this problem, the authors of this paper begin by proposing a general formulation for sparse kernel learning that is based on the random feature approximation. In this formulation, the loss functions may or may not be convex. This formulation, which is based on the orthogonal random feature approximation, is another one that we use in order to cut down on the number of random features that are necessary for an experiment. After that, we put forward an original asynchronous parallel doubly stochastic algorithm for large-scale sparse kernel learning (AsyDSSKL). As far as we are aware, the AsyDSSKL algorithm is the first one to combine the methods of asynchronous parallel computation and doubly stochastic optimization. Additionally, we offer a comprehensive convergence guarantee for the AsyDSSKL algorithm. Importantly, the experimental results on a variety of large-scale real-world datasets show that our AsyDSSKL method has a significant advantage over the existing kernel methods in terms of the computational efficiency at the training and predicting steps.