Visualizing and Understanding Policy Networks of Computer Go

In May 2017, the application of deep learning to game “Go” enjoyed a tremendous victory when AlphaGo computer program beat one top professional players. However, there is no clear understanding why elicits such strong performance. this paper, we introduce visualization techniques used in image recognition investigate functions intermediate layers and operations Go policy network. Used as diagnostic tool, these allow us understand what happens during training networks. We also technique that performs sensitivity analysis classifier output by occluding portions input board, revealing which parts board are important for predicting next move.