Thesis Defense: Samuel Tenka
4-257
Name:
Samuel Tenka
Time:
Wednesday, June 18, 4pm
Room: 4-257
Zoom: https://mit.zoom.us/j/96397133905
Password: cohobast
Title: Geometric Aspects of Optimization and Representation in Learning
Abstract: Machine learning combines methods in optimization, data-representation, and architecture. Employing differential geometric tools, I develop a predictive theory of how noise-curvature interactions during optimization affect implicit regularization; characterize featurization, optimal with respect to bilinear downstream prediction, in the regime where independence constraints as expressed probabilistic graphical models are approximately but not exactly true; and prove a universal approximation theorem for equivariant architectures. Common to these three projects is an emphasis on bilinear metric structure, which behind the scenes is an instance of a category-theoretic adjunction much like that between syntax and semantics.
Supervisor:
Joshua B. Tenenbaum, Professor of Brain and Cognitive Science
Details
- Date: Wednesday, June 18
- Time: 4:00 pm - 6:00 pm
- Category: Thesis Defense
- Location: 4-257