Crafting deep learning models for reinforcement learning and computer vision applications

This thesis Crafting Deep Learning Models for Reinforcement Learning and Computer Vision Applications focuses on designing novel and effective representation learning frameworks. There are two major aspects of our proposed approaches: neural network model architecture design and objective engineering. To demonstrate how each aspect can be maneuvered, we delve into representative applications from two important areas of studies in artificial intelligence, namely reinforcement and computer vision. In both areas, we emphasize how to manipulate abstract representations to build in strong inductive biases from the target tasks and the type of available data. We hope our examples may shed light on future endeavors in tackling problems from related areas and beyond.