PLAN-IN-ADVANCE ACTIVE LEARNING OF CLASSIFIERS Pre-labelling Selection Criteria and Connection to Optimal Design of Experiments

In many sensing problems one has limited labelled data with which to learn a classifier. The acquisition of labels may be expensive, and therefore it is desirable to develop a criterion that defines which labels would be most informative to classifier design if they could be acquired. In addition, it is desirable to determine in advance all examples for which labels are desirable, so one may plan in advance an efficient means of label acquisition (e.g., when labels are labelled manually, by visiting different points in space). Toward this goal ,we present a plan-in-advance approach to active learning of classifiers. We develop selection criteria that are directly related to loss minimization and yet do not involve use of labels. This pre-labelling selection allows the experimenter to plan in advance the visits to the selected data and follow an optimal route to save unnecessary travel. We show the connection of the developed criteria to the theory of optimal experiments and demonstrate their performance on the problem of UXO detection.