The Towers Of Saliency visualization aims to grey the black box of deep neural networks. In the use case of Pommerman, a modern recreation of Bomberman that supports AI players, our collaborators at BorealisAI have built an AI player based on a Convolutional Neural Network (CNN). CNNs are inherently difficult for humans to understand; thousands and thousands of statistical calculations are hard to interpret. This visualization aims to represent the player using a more abstract, interpretable approach.
While the AI player is playing the game, a saliency mapping technique is used. Saliency mapping generates a map of for a given input of which areas the learner finds more important in determining it’s output. In this case the learner is the AI player, and the output is the action the player thinks it should take. A map can be made for each decision the player makes, and so this visualization stacks each map on top of each other to be viewed using immersive reality technologies.