Event Date
Speaker: Tuan Pham, Post-Doctoral Fellow, University of Texas, Austin
Title: "Time-uniform bounds for iterative algorithms"
Abstract: Modern machine learning algorithms often involve making data-dependent decisions, which in turn lead to random stopping rules that are not fixed in advance.
In contrast, many existing theoretical results for these algorithms only provide guarantees at a fixed time specified in advance.
It is therefore of interest to develop theoretical guarantees that hold not just at a single time, but uniformly over the entire time horizon.
In this talk, I will discuss recent progress on such time-uniform results and present a new framework that works well for one-pass algorithms, such as stochastic gradient descent (SGD), Oja's algorithm for PCA, and the Robbins--Monro scheme.
Unlike existing approaches that rely on constructing a single “one-shot’’ supermartingale, our framework only requires understanding how the loss function changes after one iteration and is closely related to the classical Robbins--Siegmund lemma.
This is based on joint work with Alessandro Rinaldo and Purnamrita Sarkar, paper arXiv:2511.18273
This talk is part of the STA 290 Seminar series.