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