PoP Seminar Talk

Bridging the Theory and Practice of Continuous Computations

Wonyeol Lee, Postdoctoral Associate, Department of Computer Science, Carnegie Mellon University
Thursday, 01 February, 2024; 3:00pm
NSH 3305
Host: Feras Saad

Abstract

Continuous computations, which involve continuous data and operations, play a key role in many areas such as machine learning and scientific computing. In theoretical studies of these computations, we usually assume that certain mathematical objects (e.g., input functions) are well-behaved (e.g., differentiable). In practice, however, these objects are often ill-behaved (e.g., non-differentiable) partly because they are expressed in terms of programs. This gap between theory and practice hinders rigorous reasoning about these computations, and sometimes produces totally unexpected results.

In this talk, I will present some of my work on studying and bridging this gap, especially for two classes of continuous computations: automatic differentiation and variational inference. For each of these computations, I will first identify several gaps between the theory and practice of the computation (e.g., differentiable vs. non-differentiable programs), some of which have been overlooked before. I will then explain two types of new results around these gaps: how these gaps can affect the correctness of the computation, and how we can recover the correctness even under these gaps.

Bio

Wonyeol Lee is a postdoc at CMU, working with Feras Saad. He received a PhD in Computer Science from Stanford, working under Alex Aiken. In 2017-2020, he was at KAIST for military service, working with Hongseok Yang. His research aims to (i) deepen our theoretical understanding of existing practical continuous computations, and (ii) design new practical continuous computations that have theoretical guarantees.