Thursday, April 11, 2024
10:00am-12:00pm
RS208, Rosebrugh Building,
164 College St.
Justin Beland, University of Toronto
Mock Lecture (50 Mins): A Technical Introduction to Large Language Models
Technical Talk (35 Mins) & Q&A (15 Mins): Bayesian optimization under uncertainty
Optimizing the performance of complex, real-world systems under uncertainty is a challenging problem, particularly when computationally expensive predictive models are used. In this talk, I will present a Bayesian optimization approach for efficiently solving a general class of robust optimization problems involving the minimization of a specified robustness metric subject to a set of nonlinear inequality constraints. The central idea is to construct Gaussian process models to approximate robustness metrics and constraint functions. We formulate acquisition functions that leverage these Gaussian process models to identify a robust optimum solution on a limited computational budget. We also provide a theoretical upper bound on the cumulative regret of the proposed Bayesian optimization algorithm. Numerical studies on a set of test problems are presented to demonstrate the efficacy of the proposed approach.
The presentation will extend the discussion on the Bayesian optimization framework to its application in robust design. Robust design is a variant of the standard design optimization problem wherein a set of control variables have been identified (i.e., variables that are within the control of the designer) along with noise variables (i.e., variables that are inherently uncertain or expensive to control precisely but their statistics/bounds are available). The proposed research seeks to define and optimize performance measures reflecting the robustness of the design, subject to a set of deterministic/probabilistic constraints.
Bio:
Justin Beland is an instructor in the Department of Mechanical and Industrial Engineering (MIE) at the University of Toronto, specializing in machine learning. His research is conducted at the University of Toronto Institute for Aerospace Studies (UTIAS), where he investigates Bayesian optimization strategies, probabilistic modeling and deep neural network architectures.
With significant teaching experience at the University of Toronto, Justin has taught courses related to machine learning and artificial intelligence. This year marks his second time teaching APS360 Applied Fundamentals of Deep Learning, alongside other courses like MIE429 Machine Intelligence Capstone Design and MIE1517 Introduction to Deep Learning. His teaching approach is centered on making complex topics understandable and accessible, demonstrating a strong commitment to educational excellence. Additionally, Justin co-founded Analyticly Solutions, where he serves as chief scientist, contributing his expertise to the field of financial planning.
Evaluation Questions Link: https://forms.office.com/r/g9qdGbfeAV
