PhD Defense: Understanding and Enriching Algorithmic Reasoning Capabilities of Deep Learning Models

Learning to reason is an essential step to achieving general intelligence. My research has been focusing on empowering deep learning models with the abilities to generalize efficiently, extrapolate to out-of distribution data, learn under noisy labels, and make better sequential decisions --- all of these require the models to have varying levels of reasoning capabilities. As the reasoning process can be described as a step-by-step algorithmic procedure, understanding and enriching the algorithmic reasoning capabilities has drawn increasing attention in deep learning communities.To bridge algorithms and neural networks, we propose a framework, algorithmic alignment, which connects neural networks with algorithms in a novel manner and advances our understanding of how these two fields can work together to solve complex reasoning tasks. Intuitively, the algorithmic alignment framework evaluates how well a neural network's computation structure aligns with the algorithmic structure in a reasoning process. Utilizing algorithmic alignment, we are able to understand the limitations of machine learning models in the context of reasoning (e.g., generalization, extrapolation, and robustness) and know where to make improvements.Beyond this framework, we also investigate how to make better sequential decisions, during which we introduce a class of efficient approaches --- hindsight learning --- which allows us to leverage the knowledge inside existing human-designed algorithms to make better sequential decisions under uncertainty.

Examining Committee

Chair:

Dr. John Dickerson

Dean's Representative:

Dr. Mark Fuge

Members:

Dr. Jimmy Ba

Dr. Jia-Bin Huang

Dr. Petar Veličković

Dr. Tianyi Zhou