Task-free Attention Learning with Intrinsic Reward and Adversarial Learning

Abstract

Recent advances in artificial intelligence, especially deep learning, have enabled us to handle wider range of problems with computers. As for the real-world problem settings, however, there remain some difficulties, for example, inputs for embodied agents are partially observed representation of their states, and building models of their environments is needed for more sample efficient systems. One possible solution for coping with these difficulties is to use attention mechanism, which models the visual system of human and regards its inputs as a sequences, learning to where to attend. In this paper, we propose a method to train attention mechanism of neural network without external rewards. The proposed method consists of two ideas, one is to use intrinsic reward for attention mechanism and the other is to adopt adversarial learning in the model.