Dec 5 (Thu) @ 4:30pm: "Understanding Overparameterized Neural Networks: Insights into Imbalances and Neural Representations," Ganesh R Kini, ECE PhD Defense

Location: Engineering Science Bldg (ESB), Room 2003
Zoom Meeting: https://ucsb.zoom.us/j/9148373366?omn=89438978643

Abstract

This talk explores the properties of overparameterized deep neural networks, with a particular focus on training with imbalanced data. We propose a theoretically grounded loss function tailored for classification tasks under class and subgroup imbalances. Additionally, we analyze the representations learned by these networks under various training objectives, uncovering a geometrical characterization of the learned features and classifier head weights. This geometrical perspective offers insights into the influence of loss functions, hyperparameter choices, and mini-batch strategies, providing a deeper understanding of how imbalances affect network behavior and performance.

Bio

Ganesh R Kini is a PhD candidate in the Electrical and Computer Engineering department at UC Santa Barbara. He holds an MS from the Indian Institute of Science. His research interests include robust learning under imbalances, representation learning, and diffusion models for sensory data.

Hosted By: Prof Ramtin Pedarsani

Submitted By: Ganesh Ramachandra Kini <kini@ucsb.edu>