DEPS Abstract

Guest >> Sign In

DIRECTED ENERGY PROFESSIONAL SOCIETY

Abstract: 25-Symp-112

UNCLASSIFIED, PUBLIC RELEASE

Neural Collapse II: Demonstration and Robustness of Phenomenon

Neural Collapse is a phenomenon where a neural network’s trained structure converges to mathematically
well-defined properties which correlate to desirable behaviors such as performance and generalization. This talk
documents various Phase-I STTR efforts exploring Neural Collapse mechanisms. The goals of Phase-I include
validation of the phenomenon’s existence, develop an understanding of when and where it occurs, and prepare for
Phase-II, which explores the implications of Neural Collapse. Listeners will learn about the phenomenon, why it is
being researched, and how it relates to modern machine learning research and datasets. This is a continuation of
research presented at DE Systems 2024.

UNCLASSIFIED, PUBLIC RELEASE