Doctoral Thesis: Parameterizations of Neural Fields

By: Clinton Wang

Supervisor: Polina Golland

Details

• Date: Wednesday, August 7
• Time: 10:00 am - 12:00 pm
• Category: Thesis Defense
• Location: 32-G449 (Patil/Kiva)

Additional Location Details:

Abstract: Neural fields are an invaluable tool in the modern repertoire of signal representations, finding success in diverse applications across many types of signals. The first part of this talk explores the design space of neural field parameterizations, focusing on two very different tasks—novel view synthesis in large 3D scenes, and motion stabilization in volumetric time series. We describe techniques to make neural fields better matched to these signals and task requirements. Then, we describe methods for using neural fields as datapoints for data-driven learning, which addresses a key shortcoming of neural fields relative to conventional signal representations. Since the heterogeneity of neural field designs makes most data-driven approaches unusable, we introduce a sampling-based approach that is agnostic to how the field is parameterized. This method enables tasks like classification or representation learning to be performed on neural fields analogously to discrete signals like images.