Active Learning for Convolutional Neural Networks: A Core-Set Approach.

Convolutional neural networks (CNNs) have been successfully applied to manyrecognition and learning tasks using a universal recipe; training a deep modelon a very large dataset of supervised examples. However, this approach israther restrictive in practice since collecting a large set of labeled imagesis very expensive. One way to ease this problem is coming up with smart waysfor choosing images to be labelled from a very large collection (ie. activelearning).

Our empirical study suggests that many of the active learning heuristics inthe literature are not effective when applied to CNNs in batch setting.Inspired by these limitations, we define the problem of active learning ascore-set selection, ie. choosing set of points such that a model learned overthe selected subset is competitive for the remaining data points. We furtherpresent a theoretical result characterizing the performance of any selectedsubset using the geometry of the datapoints. As an active learning algorithm,we choose the subset which is expected to yield best result according to ourcharacterization. Our experiments show that the proposed method significantlyoutperforms existing approaches in image classification experiments by a largemargin.