PhD Thesis Defense - Muchen Li

Name: Muchen Li

Date: Friday, October 9, 2026

Time: 10:00 AM

Location: Room 146, ICICS, 2366 Main Mall

Zoom: https://ubc.zoom.us/j/62538230754?pwd=lM0RXs8RWALaii3pq4byzoZ3yEwXVG.1

Supervisors' names: Professor Leonid Sigal and Professor Renjie Liao

Thesis title: Deep Generative Models for Structured Outputs: Graphs, Motions and Language

Abstract:
Generative AI models are increasingly asked to produce more than text or images: they design neural networks, locate objects described in language, synthesize how hands grasp and move objects, and draw on large datasets and memory to reason. These outputs have internal structure, and this thesis asks a practical question: where should that structure be built into a model so that it works well, generalizes to new situations, and can be reliably evaluated? I explore three strategies. The first builds structure directly into what the model outputs, for example by representing network designs as graphs. The second breaks generation into simpler, aligned parts, for example by separating how a hand approaches an object from how it grasps it, which lets the model handle objects it has never seen. The third builds structure into the data and systems behind the model, such as a large math dataset built from online community discussions and a memory mechanism that retrieves only the relevant information for each input. Together, these projects offer a comparative picture of when and how explicit structure makes generative models more capable, efficient, and trustworthy.