Dr. Andrew Ilyas, CMU: Predicting and Optimizing the Behavior of ML Models

​EvenUp is excited to host the Virtual EvenUp Distinguished Speaker Series featuring innovators, researchers and engineers in GenAI. These events will be hosted on Zoom to allow for anyone to join and learn!

For our second speaker, we are honored to host Dr. Andrew Ilyas, CMU to discuss Predicting and Optimizing the Behavior of ML Models.

Dr. Andrew Ilyas, who earned his Ph.D. from MIT and will soon join the faculty at Carnegie Mellon University, has made pioneering contributions to understanding how data influences model behavior. These insights are reshaping the way large language models are trained and fine-tuned.

In this talk, we will discuss the the problem of predicting and optimizing the counterfactual behavior of large-scale ML models.

* We will start by focusing on “data counterfactuals,” where the goal is to estimate the effect of modifying a training dataset on the resulting machine learning outputs (and conversely, to design datasets that induce specific desired behavior).

* A method will be introduced that almost perfectly estimates such counterfactuals as well as many others. This method unlocks some possibilities in the design and evaluation of ML models, including state-of-the-art data selection, attribution, and poisoning.

This is all based on work in progress with Logan Engstrom, Benjamin Chen, Axel Feldmann, Billy Moses, and Aleksander Madry.