Explanation-Based Active Learning

2018 - presentpresent

We will develop efficient algorithms and performance bounds for active learning with feature feedback. We propose to collect the feedback with the NEXT system. For example, NEXT could be used to ask people to categorize articles based on sentiment and highlight a few words that helped in making the decision. This information can then be used to improve the performance of active learning algorithms. To incorporate the information we propose to restrict the active learning algorithms only on the set of important features leading to interesting tradeoffs.