Thursday / CS Seminar

October 3rd / 12:00 PM

Hybrid / Mudd 3514

Speaker

Kawin Ethayarajh

Talk Title

Behavior-Bound Machine Learning

Abstract

We have long accepted that machine learning is bottlenecked by what hardware and software can do; we often discuss whether a process is memory-bound or compute-bound, for example. In this talk, I propose that machine learning is also behavior-bound: i.e., what it can accomplish is often circumscribed by the behavior of real-world actors such as workers, firms, and states. The behavior of these actors is only partially observable, hard to change, and often conflicts with that of other actors. My work formalizes these behaviors so that we can build machine learning pipelines that work as well in the real world as they do on paper, with actual humans and not just idealized ones. In this talk, I will discuss: (1) how to overcome the principal-agent problem in dataset creation to create datasets as complex as real-world problems; (2) how Kahneman & Tversky's prospect theory enables us to better align models with human feedback; (3) how to do cost-sensitive evaluation that better captures how firms choose among models in practice.

Biography

Kawin Ethayarajh is a Postdoctoral Fellow at Princeton Language and Intelligence (PLI), where he works on behavioral machine learning. He received a PhD in Computer Science from Stanford University, where he was advised by Dan Jurafsky. Prior to that, he was a National Scholar at the University of Toronto, where he received a BSc and MSc in Computer Science. His work has been recognized with a Facebook (Meta) Fellowship and an Outstanding Paper award at ICML 2022.

---

Zoom: https://northwestern.zoom.us/j/96280259503

Panopto: https://northwestern.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=c29e2db7-f445-4437-9290-b1ef00f262af

DEI Minute: Colorblind Accessibility https://chemistry.northwestern.edu/documents/colorblind-accessibility.pdf