About me
I am a final-year Ph.D. candidate in Computer Science at UT Austin. I am fortunate to be advised by Prof. Adam Klivans. Before that, I studied Electrical and Computer Engineering at the National Technical University of Athens, where I worked with Prof. Dimitris Fotakis. I am grateful to be supported by the 2025 Apple Scholars in AI/ML Fellowship. In summer 2025, I was a research intern at Apple MLR, working with Parikshit Gopalan and Kunal Talwar.
My research focuses on the computational foundations of reliability in machine learning. I am particularly interested in designing efficient learning algorithms with provable guarantees that do not rely on strong distributional or modeling assumptions, especially in challenging scenarios like learning under distribution shift or data contamination.
News
Publications
Learning with Distribution Shift
Classical learning formulations assume train and test data come from the same distribution, which rarely holds in practice. I design efficient algorithms that provably succeed under distribution shift, including tolerant and testable guarantees that certify when their output can be trusted.
Selected papers (7)
Learning from Noisy or Corrupted Data
Training datasets are often contaminated by naturally occurring or adversarial corruptions. I develop efficient algorithms that learn with provable guarantees under challenging noise models, ranging from Massart noise and adversarial label noise to nasty noise and heavy additive contamination.
Selected papers (4)
Testable Learning
The testable learning framework asks an algorithm to verify, rather than assume, that its distributional assumptions hold. I build efficient tester-learners that either return a classifier with a certificate of near-optimality or reject the data.
Selected papers (3)
Beyond Worst-Case Models
Worst-case analysis can be overly pessimistic about what is learnable. I use beyond-worst-case lenses, such as smoothed or average-case analysis, to explain why learning remains tractable on realistic, mildly perturbed instances.
Selected papers (2)
Calibration and Real-Valued Prediction
Beyond binary classification, reliable predictions require well-calibrated probabilities and accurate real-valued outputs. I study how to verify or achieve such properties efficiently, with applications to decision making and omniprediction.