Active Sensing

Labels are often expensive to get, and this motivates active learning which chooses the most informative samples for label acquisition. In this paper we study active sensing in a multi-view setting, motivated from many problems where grouped features are also expensive to obtain and need to be acquired (or sensed) actively (e.g., in cancer diagnosis each patient might go through many tests such as CT, Ultrasound and MRI to get valuable features). The strength of this model is that one actively sensed (sample, view) pair would improve the joint multi-view classification on all the samples. For this purpose we extend the Bayesian co-training framework such that it can handle missing views in a principled way, and introduce two criteria for view acquisition. Experiments on one toy data and two real-world medical problems show the effectiveness of this model.