Lifelong Learning with Output Kernels

Lifelong learning poses considerable challenges in terms of effectiveness (minimizing prediction errors for all tasks) and overall computational tractability for real-time performance. This paper addresses continuous lifelong multitask learning by jointly re-estimating the inter-task relations (\textit{output} kernel) and the per-task model parameters at each round, assuming data arrives in a streaming fashion. We propose a novel algorithm called \textit{Online Output Kernel Learning Algorithm} (OOKLA) for lifelong learning setting. To avoid the memory explosion, we propose a robust budget-limited versions of the proposed algorithm that efficiently utilize the relationship between the tasks to bound the total number of representative examples in the support set. In addition, we propose a two-stage budgeted scheme for efficiently tackling the task-specific budget constraints in lifelong learning. Our empirical results over three datasets indicate superior AUC performance for OOKLA and its budget-limited cousins over strong baselines.