Interactive machine learning

Interactive machine learning is a paradigm to actively acquire information, and minimize unnecessary interactions with an expert. It characterizes a broad range of decision making problems arising in pattern recognition and related fields. It generalizes tasks such as Bayesian experimental design, where one needs to adaptively pick a small sets of experiments in order to estimate uncertain model parameters. For these problems, we recently introduced a structural property, adaptive submodularity, which generalizes the classical notion of submodular functions to adaptive policies. We show that if a problem satisfies this property, an efficient, adaptive greedy algorithm guarantees near-optimal policies. We further show that this property arises in diverse applications including active learning, adaptive viral marketing, active object detection, and optimizing value of information.