Visual Explanations in Deep Reinforcement Learning via Convolutional Neural Network Localization

Deep Reinforcement Learning (DRL) connects the classic Reinforcement Learning algorithms with Convolutional Neuronal Networks (CNN). The problem in DRL is that it is hard to understand what the CNN is learning. In order to be able to use the programs in highly dangerous environments for humans and machines, the developer need a debugging tool to improve that the program does what he expects.
In this presentation I will show and explain how guided backpropagation algorithm can help to debug algorithms. Beside this I will introduce a new architecture of an off policy model which combines the attention network with the dueling network.