Optimal Learning With Conditional Guarantees Georgy Noarov

Format:

Book

Thesis/Dissertation

Author/Creator:

Noarov, Georgy, author.

Contributor:

University of Pennsylvania. Computer and Information Science., degree granting institution.

Language:

English

Subjects (All):

0984.

0489.

0800.

0463.

Local Subjects:

0984.

0489.

0800.

0463.

Genre:

Academic theses

Physical Description:

1 electronic resource (556 pages)

Contained In:

Dissertations Abstracts International 88-03B

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2026

Language Note:

English

Summary:

Modern ML and AI systems have achieved remarkable performance across diverse domains. However, their growing predictive power and flexibility have been accompanied by persistent concerns about trustworthiness and robustness. Given the usual emphasis on optimizing expected loss metrics, a model may perform excellently on average while lacking guarantees on small but relevant slices of the data such as decision-critical subpopulations, limiting the reliability of its predictions for downstream decision making. This motivates the development of methods that endow ML models with provable conditional validity. In this dissertation, we approach this problem via the lens of fine-grained uncertainty quantification. The overarching goal is to develop a toolkit for training or post-processing ML models so that their predictions pass rich families of conditional validity tests, behaving reliably not only on average but also on relevant groups of examples and thus inducing high-quality and trustworthy decisions. Specifically, we develop flexible frameworks and computationally efficient algorithms for learning predictors with provable conditional calibration guarantees. These frameworks apply even in online adversarial settings, where data arrive sequentially and may be adversarial generated. On the sample-complexity side, we obtain minimax error rates for online multicalibration and related notions, giving optimal-rate algorithms together with matching or near-matching lower bounds. Further, we rigorously demonstrate a range of benefits that our toolkit yields in a variety of settings. Thus, we develop the first scalable and general conditional conformal prediction methods, in the batch and online adversarial settings. We also formulate a general event-conditional prediction framework that enables high-dimensional sequential decision making with provable no-regret guarantees. Applying our ideas to causal inference, we design multigroup adaptive RCTs with provable ATE estimation guarantees. Last but not least, we study calibration in the context of LLMs, and in particular develop rigorous, calibration-inspired foundations for top-k decoding. In conclusion, we provide both foundations and numerous applications of fine-grained calibration methods with conditional guarantees

Notes:

Source: Dissertations Abstracts International, Volume: 88-03, Section: B.

Advisors: Kearns, Michael; Roth, Aaron Committee members: Vohra, Rakesh; Dobriban, Edgar; Hassani, Hamed; Kleinberg, Robert

Ph.D. University of Pennsylvania 2026

Includes bibliographical references

Local Notes:

School code: 0175

ISBN:

9798189294174

Access Restriction:

Restricted for use by site license