Written by Sofia Hedegaard Rasmussen
Machine learning is rapidly becoming part of the technologies we rely on every day. From the tools we use at work to the systems shaping healthcare, finance, and digital services, machine learning is increasingly influencing how decisions are made.
But what happens when the data behind these systems is flawed, manipulated or simply doesn't behave as expected?
That is one of the questions that drives Arsen Vasilyan's research as he joined the department as a new Assistant Professor on 1 August.
"Machine learning tools are being deployed everywhere right now. They are reshaping how we do everything, and our societies are becoming increasingly dependent on them. Yet, these tools still have significant reliability issues and vulnerabilities that nobody knows how to overcome," he says.
Arsen's research lies in computational learning theory and the foundations of machine learning. Rather than developing specific AI applications, he studies the theoretical principles that determine when and why machine learning algorithms can be trusted, and how to make them more robust when things go wrong.
Machine learning systems are often remarkably capable, but they can also be surprisingly trusting. Most algorithms assume that the data they are trained on is accurate and representative. However, data can be noisy, incomplete or even intentionally manipulated.
Arsen develops algorithms that continue to perform reliably even when these assumptions break down. For example, some of his work investigates how machine learning models can learn successfully even when part of their training data has been deliberately altered by a malicious actor attempting to make the system fail.
"It is impossible to know in advance how bad actors will behave or what type of strategy they will use. That is why these settings require a good theoretical understanding and formal reliability guarantees," he explains.
His research also explores how algorithms can recognise when they should not trust the data they are given. Rather than producing potentially unreliable predictions, future machine learning systems may be able to detect when their underlying assumptions no longer hold and alert users that they cannot provide a trustworthy answer.
Machine learning has seen spectacular progress in recent years. Yet our theoretical understanding of why these systems work, and when they fail, remains limited.
"Machine learning has been spectacularly successful in recent years, but the theoretical understanding is still lagging far behind," Arsen says.
This gap between theory and practice presents one of the biggest challenges in the field. At the same time, it represents an opportunity to shape the future development of reliable and trustworthy machine learning.
Arsen is particularly excited about narrowing this gap by using theoretical insights to develop machine learning algorithms that are both more robust and more practical. His recent work includes developing new methods for learning fundamental machine learning models in the presence of corrupted data and designing algorithms that can test whether the assumptions behind a learning algorithm hold before providing an answer.
Born and raised in Armenia, Arsen's academic journey has taken him across some of the world's leading institutions in theoretical computer science and machine learning.
Before joining Aarhus University, he was a postdoctoral fellow at the Institute for Foundations of Machine Learning at the University of Texas at Austin and a research fellow at the Simons Institute for the Theory of Computing at UC Berkeley. He completed his PhD in Computer Science at the Massachusetts Institute of Technology (MIT), where he was advised by Ronitt Rubinfeld and Jonathan Kelner.
Now, Aarhus is the next stop on that journey.
What attracted Arsen to the department was its strong theoretical computer science environment and collaborative culture.
"Aarhus is a fantastic place to work, collaborate, and advise students and postdocs. I think it is important for students in my field to be exposed to ideas and approaches from many areas of theoretical computer science, and Aarhus is a great place for this because it has so many excellent researchers," he says.
He particularly appreciates that students are encouraged to collaborate broadly rather than limiting themselves to a single research area or supervisor. Having benefited from such an environment himself as a student, he hopes to provide similar opportunities for his own students and collaborators. Looking ahead, Arsen would like to teach courses that introduce students to computational learning theory and the foundations of machine learning. He also looks forward to supervising Master's and PhD students and welcoming new researchers into the field.
"I am always happy to chat about our research and introduce new people to our research field," he says.
When he is not thinking about machine learning, Arsen enjoys hiking, exercising, cooking and improv comedy.
A Fully Polynomial-Time Algorithm for Robustly Learning Halfspaces over the Hypercube
Gautam, C., Klivans, A.R., Stavropoulos, K, & Vasilyan, A.
58th ACM Symposium on Theory of Computing (STOC 2026)
Halfspaces (also known as linear classifiers) are one of the most basic types of hypotheses. They are one of the first things we teach in an introductory machine learning task. Yet, we do not fully understand how to learn halfspace if the training data has noise and corruptions. In this work we make progress in this research direction.
Learning Constant-Depth Circuits in Malicious Noise Models
Klivans, A. R., Stavropoulos K. & Vasilyan, A.
Published at 38th Conference on Learning Theory (COLT 2025)
In this work we show how to learn another very basic class of functions in the presence of noise and corruptions in the training data. Specifically, we study the class of constant-depth circuits. We develop a new method to filter the training data that allows our learning algorithm to succeed in the presence of data corruption.
Testing Distributional Assumptions of Learning Algorithms
Rubinfeld, R. & Vasilyan, A.
55th ACM Symposium on Theory of Computing (STOC 2023)
A lot of times the reliability of machine learning algorithms is based on various assumptions about their training data. If such an assumption does not hold, the reliability guarantees will fail. In this paper we give algorithms that will alert users that the assumption is wrong instead of giving an unreliable answer. Since our initial paper, there has been a lot of follow-up work by us and other researchers that develop such algorithms in many other settings.
