archimedes-Artificial Intelligence, Data Science, Algorithms-greece

[Archimedes Talks Series]Efficiently Certifiable Guarantees for Learning with Distribution Shift

Dates
2025-01-08 14:30 - 16:00
Venue
Artemidos 1 - Amphitheater
TITLE: Efficiently Certifiable Guarantees for Learning with Distribution Shift

SPEAKER: Kostas Stavropoulos (Ph.D. student in Computer Science at UT Austin)

ABSTRACT: Learning in the presence of distribution shift remains a major and challenging problem in machine learning. In this setting, the learner is trained on some labeled training distribution, but evaluated on some other, potentially adversarial, test distribution for which the learner only has unlabeled examples. A long series of works in the past twenty years has focused on giving bounds for the test error in terms of appropriate notions of distance between the training and test distributions. Such distances, however, typically involve enumerations and no efficient algorithms for estimating or even testing such distances are available.

In this talk, we present a new model called testable learning with distribution shift, where the learner is allowed to reject, but only if distribution shift is detected. If the learner accepts, then the output hypothesis is guaranteed to have low test error. In this framework, we provide the first efficient algorithms for learning several fundamental concept classes in the presence of distribution shift under standard assumptions on the training marginal distribution. The classes we capture include halfspace intersections, decision trees and, in general, any class that admits low-degree sandwiching approximators.

REFERENCES: https://arxiv.org/abs/2311.15142 , https://arxiv.org/abs/2404.02364 , https://arxiv.org/abs/2406.09373 , https://arxiv.org/abs/2406.02742 

SHORT BIO: Kostas Stavropoulos ( https://www.kstavrop.com ) is currently a fourth-year Ph.D. student in Computer Science at UT Austin. He is fortunate to be advised by Prof. Adam Klivans. Before that, he studied Electrical and Computer Engineering at the National Technical University of Athens, where he was fortunate to work with Prof. Dimitris Fotakis. His research is in the intersection of machine learning and theoretical computer science. He is particularly interested in designing efficient learning algorithms with provable guarantees that do not rely on the strong assumptions typically made in learning theory, especially in challenging scenarios like learning with distribution shift and/or noise.


________________________________________________________________________________
Microsoft Teams Need help?
Join the meeting now
Meeting ID: 364 928 182 762
Passcode: w24fHy
________________________________________________________________________________