Active and Continuous Exploration with Deep Neural Networks and Expected Model Output Changes

The demands on visual recognition systems do not end with the complexity by current large-scale image datasets, such as ImageNet. In, we need curious and continuously learning algorithms that actively knowledge about semantic concepts which are present in available data. As a step towards this goal, we show how to perform continuous learning and exploration, where an algorithm actively selects relevant of unlabeled examples for annotation. These examples could either to already known or to yet undiscovered classes. Our algorithm is based a new generalization of the Expected Model Output Change principle for deep and is especially tailored to deep neural networks. Furthermore, show easy-to-implement approximations that yield efficient techniques for selection. Empirical experiments show that our method outperforms used heuristics.