I am a Postdoctoral Fellow in the Harvard John A. Paulson School of Engineering and Applied Sciences at Harvard University, hosted by Sitan Chen. Before that, I was an FDS Postdoctoral Fellow at Yale University. I obtained my PhD in the Computer Science Department of the National Technical University of Athens (NTUA) working with Dimitris Fotakis and Christos Tzamos.
I am interested in the statistical and computational foundations of machine learning, especially their applications to causal inference and generative modeling as well as their connections to theoretical computer science and statistical physics.
My research studies the statistical and computational foundations of machine learning. I am currently mostly interested in their applications to causal inference, generative modeling, and the study of computation in physical systems. Below there is an overview of these research directions, along with some representative papers.
Missing and selectively observed data arise throughout science in fields like biology, economics, and the social sciences, inducing systematic bias in the statistical pipeline, leading to incorrect conclusions and poor decision-making.
Generative models have amazing capabilities that span a wide range of applications including image and code generation, as well as mathematical reasoning. Their empirical success therefore raises a fundamental theoretical question:
Statistical physics models provide a rich setting for studying computational questions arising in machine learning, including sampling and inference.
Learning theory seeks principles that characterize statistical and computational complexity across learning tasks.
Some earlier research directions include the questions of (i) what does it mean for an algorithm to be replicable? What is the cost of replicability as a property of the algorithm? and (ii) what is the complexity of reaching equilibria in multi-agent settings?
Instructor: Alkis Kalavasis
This course is about generalization and stability of Machine Learning (ML) systems. There are various ways to define what it means for a learning algorithm to be stable. The most standard way is inspired by sensitivity analysis, which aims at determining how much the variation of the input can influence the output of a system. This abstract way allows one to introduce various notions of stability such as uniform stability, differential privacy, and replicability. In this course, we investigate these notions of stability, their implications to learning theory, and their surprising connections.
Lecture Notes (PDF)with Andrew Ilyas, Anay Mehrotra, and Manolis Zampetakis
Area Chair: NeurIPS (2026)
Reviewer: FOCS (2026, 2025, 2024), STOC (2026, 2025, 2024), COLT (2026, 2025), NeurIPS (2024, 2023, 2022, 2021), ICML (2023), AISTATS (2022, 2021), ICLR (2022), ITCS (2024)
