Timezone: »

Machine versus Human Attention in Deep Reinforcement Learning Tasks
Sihang Guo · Ruohan Zhang · Bo Liu · Yifeng Zhu · Dana Ballard · Mary Hayhoe · Peter Stone

Tue Dec 07 04:30 PM -- 06:00 PM (PST) @

Deep reinforcement learning (RL) algorithms are powerful tools for solving visuomotor decision tasks. However, the trained models are often difficult to interpret, because they are represented as end-to-end deep neural networks. In this paper, we shed light on the inner workings of such trained models by analyzing the pixels that they attend to during task execution, and comparing them with the pixels attended to by humans executing the same tasks. To this end, we investigate the following two questions that, to the best of our knowledge, have not been previously studied. 1) How similar are the visual representations learned by RL agents and humans when performing the same task? and, 2) How do similarities and differences in these learned representations explain RL agents' performance on these tasks? Specifically, we compare the saliency maps of RL agents against visual attention models of human experts when learning to play Atari games. Further, we analyze how hyperparameters of the deep RL algorithm affect the learned representations and saliency maps of the trained agents. The insights provided have the potential to inform novel algorithms for closing the performance gap between human experts and RL agents.

Author Information

Sihang Guo (University of Texas at Austin)
Ruohan Zhang (Stanford University)
Bo Liu (Stanford University)
Yifeng Zhu (The University of Texas at Austin)
Dana Ballard (University of Texas, Austin)

Dana H. Ballard obtained his undergraduate degree in Aeronautics and Astronautics from M.I.T. in 1967. Subsequently he obtained MS and PhD degrees in information engineering from the University of Michigan and the University of California at Irvine in 1969 and 1974 respectively. He is the author of two books, Computer Vision (with Christopher Brown) and An Introduction to Natural Computation. His main research interest is in computational theories of the brain with emphasis on human vision. His research places emphasis on Embodied Cognition. Starting in 1985, he and Chris Brown designed and built the first high-speed binocular camera control system capable of simulating human eye movements in real time. Currently he pursues this research at the University of Texas at Austin by using model humans in virtual reality environments. His focus is on the use of machine learning as a model for human behavior with an emphasis on reinforcement learning

Mary Hayhoe (University of Texas, Austin)
Peter Stone (The University of Texas at Austin, Sony AI)

More from the Same Authors