Artificial intelligence is beyond human intelligence. This thesis is about modeling intelligence that can learn to represent and reason about the world. We study both questions from the lens of graph neural networks. First, we can abstract many objects in the world as graphs and learn their representations with graph neural networks. Second, learning to reason implies learning to implement a correct reasoning process, within and outside the training distribution. We shall see how graph neural networks exploit the algorithmic structure in reasoning processes to improve the generalization.
Each part of the thesis also studies one aspect of the theoretical landscape of learning: the representation power, generalization, extrapolation, and optimization.
I: We build powerful graph neural networks.
II: We analyze generalization by considering the training algorithm, network structure, and task structure.
III: We study how neural networks extrapolate and show implications for learning reasoning outside the training distribution.
IV: We show global convergence rates and practically accelerate the training.
Thesis Supervisor: Prof. Stefanie Jegelka
To attend this defense, please contact the doctoral candidate at keyulu at mit dot edu