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

Abstract

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