Algorithm Selection for Deep Active Learning with Imbalanced Datasets

Active learning aims to reduce the number of labeled examples needed to train deep networks, but the empirical performance of active learning algorithms can vary dramatically across datasets and applications.