Colloquium Details
Learning Theoretic Foundations for Modern (Data) Science
Speaker: Allen Liu, MIT
Location: 60 Fifth Avenue 150
Date: February 11, 2025, 2 p.m.
Host: Richard Cole
Synopsis:
In this talk, I will explain how fundamental problems in computational learning theory are at the heart of modern problems in machine learning and scientific applications and how algorithmic insights in mathematically tractable models can inspire new solutions in a wide variety of domains.
I will explore two directions. First, I will explore algorithmic foundations for model stealing of language models. Model stealing, where a learner tries to recover an unknown model through query access, is a critical problem in machine learning. Here, I will aim to build a theoretical foundation for designing model stealing algorithms. Second, I will introduce Hamiltonian learning, a central computational task towards understanding and benchmarking quantum systems. I will highlight how the lens of learning theory plays a key role in identifying and circumventing previous barriers and allows us to give efficient algorithms in settings that were previously conjectured to be intractable.
Speaker Bio: