Active learning for black-box models

Abstract

Active learning refers to the settings in which a machine learning algorithm (learner) is able to select data from which it learns, and by doing so aims to achieve a better accuracy (e.g. by avoiding obtaining training data that is redundant or unimportant). Active learning is particularly useful in cases where obtaining training data is costly. A common assumption is that an active learning algorithm is fully aware of the details of an underlying learning algorithm for which it obtains the data. However, in many real world settings, obtaining precise details of the learning algorithm may not be feasible, making the underlying algorithm in essense a black box – no knowledge of internal workings of the algorithm, where only the inputs and corresponding outputs are acessible. We note that accuracy will improve only if the learner’s outputs change. Motivated by this, we select a training point that is expected to cause many changes in the learner’s outputs, in the anticipation that the resulting changes will be for the better.