All Times EDT
Active learning focuses on minimizing the number of labels needed to perform a task by using feedback to sequentially collect the most informative labels. Often the feedback is collected from humans who are better at making comparative judgements than providing direct labels. In this talk, I will present an interactive learning framework where the algorithm adaptively combines few (expensive and more time consuming) direct labels with (cheaper and quicker) indirect pairwise comparisons between data points, automatically deciding which type of data (label vs comparisons) to collect, when and how much. We will examine settings where access to comparisons can significantly improve label complexity in several learning tasks including classification, regression and bandit optimization, even when the comparisons may come from a different but related function than the labels.