Prof. Pascal Poupart will present techniques to automatically infer high-level object information from low-level pixel intensities to accelerate the optimization of a good policy in deep reinforcement learning.
This talk will cover the approach, named MOREL (Motion Oriented REinforcement Learning), which consists of a self-supervised technique that learns to detect moving objects in video games.
This technique allow RL agents to reason at a high level and therefore need less interaction with the environment to find good policies. Furthermore, we can gain insights and some degree of explainability into the resulting policies by inspecting the objects they depend on. The techniques will be demonstrated in Atari Games and TextWorld.