Active Information Acquisition

摘要: We propose a general framework for sequential and dynamic acquisition of useful information in order to solve particular task. While our goal could principle be tackled by reinforcement learning, setting is constrained enough allow more efficient algorithms. In this paper, we work under the Learning Search show how formulate finding policy that framework. apply formulation on two tasks, sentiment analysis image recognition, learned policies exhibit good statistical performance. As an emergent byproduct, tendency focus most prominent parts each instance give harder instances attention without explicitly being trained do so.