Geometric Concept Acquisition by Deep Reinforcement Learning

Explaining how intelligent systems come to embody knowledge of deductive concepts through inductive learning is a fundamental challenge of both cognitive science and AI. This project address this challenge by exploring how deep reinforcement learning agents, occupying settings similar to early-stage mathematical concept learners come to represent ideas such as rotation and translation. I first train a Dueling Deep Q-Network on a shape sorting task requiring implicit knowledge of geometric properties, then transfer its learned parameters to tasks querying explicit geometric knowledge, such as classification. Furthermore I conduct a number of qualitative experiments that offer intuitions as to how shape identity is encoded in the network.