KERNEL-BASED EFFICIENT LIFELONG LEARNING ALGORITHM

Multitask learning leverages shared structure across multiple tasks to obtain classifiers with generalization capability surpassing that of independent single task learning. Lifelong learning further tackles the challenge of performing multitask learning in an online fashion for a continual stream of tasks. In this work, kernel-based lifelong learning algorithm is proposed to capture significant nonlinear structure in the data. It is postulated that the classifiers accommodate a union-of-subspace model in the feature space. A shared library of atoms are then learned based on online kernel dictionary learning in a reproducing kernel Hilbert space. To alleviate the inherent complexity of nonparametric learning which grows with the data set size, an approximate classifiers are also obtained, which are representable using a parsimonious pool of selected examples. Preliminary tests verify the efficacy of the proposed methods.