Learning by Descent: A Geometric View of Optimization in Machine Learning

Date: Thursday, January 29, 2026

Time:  11:45 a.m. to 12:45 p.m.

Location: G01 Gates Hall


Abstract: Optimization via gradient descent underlies many of the most successful machine learning models in use today. This talk presents a geometric perspective on gradient descent that highlights how local information encoded by the gradient interacts with the global structure of a loss function. By studying simple but representative loss landscapes, we analyze convergence behavior, instability due to step size, and common limitations encountered in practice. The lecture emphasizes intuition and visualization, preparing students to better understand and reason about more advanced variants such as stochastic and adaptive methods.