10/31/2024
By Jack Shannon
Candidate Name: Jack Shannon
Degree: Master's
Defense Date: Thursday, Nov. 14, 2024
Time: 10 a.m. to Noon
Location: Olney Hall, Room 430, North Campus
Thesis Title: Thermodynamic Learning and Computing of Generative Stochastic Artificial Neural Networks
Advisor:
Caroline Desgranges, Research Assistant Professor, University of Massachusetts Lowell
Committee Members:
Hugo Ribeiro, Assistant Professor, University of Massachusetts Lowell
Viktor Podolskiy, Professor, University of Massachusetts Lowell
Brief Abstract:
The application of statistical physics in machine learning was a fundamental stepping stone in the advancements made for modern machine learning and AI. Stochastic neural networks such as the Restricted Boltzmann machine (RBM) act as a crucial intersection between the two fields of study. These models use statistical thermodynamic principles to be able to learn any probability distribution. However, understanding exactly how machine learning models learn in general is still unknown. The motivation behind this work is to achieve a better understanding of the physical principles governing machine learning models by observing how the RBM samples the model distribution under a variety of different temperature and non-isothermal environments, while also utilizing different learning objectives. In the end, this enables a broader picture of what the physics governing these models are doing. This will inevitably allow for more efficient, accurate, and interpretable models to be made in the future.