Abstract:
Active learning improves the efficiency of machine learning in situations where labels are acquired at a cost. In this paper, I explore how active learning techniques can be used in a lifelong machine learning setting, where the agents face a pool of tasks and acquire knowledge incrementally by learning a sequence of tasks. I develop two active learning frameworks for Efficient Lifelong Learning Algorithm (ELLA). The first, ELLA-ATAL is a flexible framework that combines active task selection and active instance selection and bridges lifelong learning algorithms with traditional single-task active learning methods. The second, ELLA-ATMIS enhances ELLA-ATAL by taking into account how each instance contributes to tasks that it does not belong to. I evaluate these approaches on two data sets and compare them with other lifelong learning methods. My results demonstrate the potential of active learning when used in combination with lifelong learning.