Doctoral Thesis: From Data, to Models, and Back: Making ML “Predictably Reliable”

Friday, August 23
2:30 pm - 4:00 pm

Kiva (32-G449)

Add to Calendar
FacebookTwitterLinkedInEmailShare

By: Andrew Ilyas

Thesis Supervisors: Costis Daskalakis, Aleksander Madry

Details

  • Date: Friday, August 23
  • Time: 2:30 pm - 4:00 pm
  • Category:
  • Location: Kiva (32-G449)
Additional Location Details:

Abstract: Despite their impressive performance, training and deploying ML models is currently a somewhat messy affair. But does it have to be? In this defense, I’ll discuss some of my research on making ML “predictably reliable”—enabling developers to know when their models will work, when they will fail, and why. To begin, we use a case study of adversarial examples to show that human intuition can be a poor predictor of how ML models operate. Motivated by this, we present a few lines of work that aim to develop a precise understanding of the entire ML pipeline: from how we source data, to the datasets we train on, to the learning algorithms to use.

Host