Predicting and Optimizing the Behavior of Large ML Models

This talk by Andrew Ilyas from Stanford University explores the challenge of predicting and optimizing counterfactual behavior in large-scale machine learning models. Discover methods for estimating how modifications to training datasets affect ML outputs, and learn techniques for designing datasets that induce specific desired behaviors. Examine an innovative approach that nearly perfectly estimates data counterfactuals, opening new possibilities in ML model design and evaluation. The presentation covers state-of-the-art methods for data attribution, selection, and poisoning as part of the "The Future of Language Models and Transformers" series at the Simons Institute.