Selective eye-gaze augmentation to enhance imitation learning in Atari games

This paper presents the selective use of eye-gaze information in learning human actions Atari games. Extensive evidence suggests that our eye movements convey a wealth about direction attention and mental states encode necessary to complete task. Based on this evidence, we hypothesize eye-gaze, as clue for direction, will enhance from demonstration. For purpose, propose augmentation (SEA) network learns when information. The proposed architecture consists three sub-networks: gaze prediction, gating, action prediction network. Using prior 4 game frames, map is predicted by network, which used augmenting input frame. gating determine whether should be fed final predict at current To validate approach, publicly available Human Eye-Tracking And Demonstration (Atari-HEAD) dataset 20 games with 28 million demonstrations 328 eye-gazes (over frames) collected four subjects. We demonstrate efficacy compared state-of-the-art Attention Guided Imitation Learning (AGIL) Behavior Cloning (BC). results indicate approach (the SEA network) performs significantly better than AGIL BC. Moreover, significance through compare random selection gaze. Even case, better, validating advantage selectively using demonstration learning.