Skip to main content

Data Science, Machine Learning, and Information Theory

Computer Science


1 Career

  • Since 2025 Professor, Institute of Science and Technology Austria (ISTA)
  • 2019 – 2025 Assistant Professor, Institute of Science and Technology Austria (ISTA)
  • 2017 – 2019 Postdoc, Stanford University, Stanford, USA
  • 2018 Research Fellow, Simons Institute for the Theory of Computing, Berkeley, USA
  • 2016 PhD, EPFL, Lausanne, Switzerland

1.1 Selected Distinctions


1.2 Selected Grants

  • 2024 – 2029 ERC Starting Grant (101161364), Inference in High Dimensions: Light-speed Algorithms and Information Limits (INF 2), Personal grant. Awarded: € 1,664,200.
  • 2024 – 2029 FWF Cluster of Excellence (10.55776/COE12), Bilateral Artificial Intelligence (BILAI), Coordinator: S. Hochreiter (JKU Linz). Awarded overall: € 33,000,000; FWF funded amount: € 19,760,512.40; awarded to RG: € 612,317.94.
  • 2019 – 2024 Lopez-Loreta Prize, Foundations of Deep Learning, Personal grant. Awarded: € 1,000,000.

1.3 Tech Transfer (Patents, Spin-Offs)

N/A


1.4 Selected Publications

  1. S. Bombari and M. Mondelli, “Privacy for free in the overparameterized regime,” Proceedings of the National Academy of Sciences, vol. 122, no. 15, e2423072122, 2025.
  2. S. Bombari and M. Mondelli, “Spurious correlations in high dimensional regression: The roles of regularization, simplicity bias and over-parameterization,” in International Conference on Machine Learning, (Best paper award at the Workshop on Spurious Correlation and Shortcut Learning, International Conference on Learning Representations, 2025), PMLR, 2025, pp. 4839–4873.
  3. S. Bombari, S. Kiyani, and M. Mondelli, “Beyond the universal law of robustness: Sharper laws for random features and neural tangent kernels,” in International Conference on Machine Learning, (Oral, top 2.4% of submissions), PMLR, 2023, pp. 2738–2776.
  4. J. Barbier, F. Camilli, M. Mondelli, and M. Sáenz, “Fundamental limits in structured principal component analysis and how to reach them,” Proceedings of the National Academy of Sciences, vol. 120, no. 30, e2302028120, 2023.
  5. A. Javanmard, M. Mondelli, and A. Montanari, “Analysis of a two-layer neural network via displacement convexity,” Annals of Statistics, vol. 48, no. 6, 2020.

2 Group Members

  • Simone Bombari, PhD Student since 2020
  • Al Depope, PhD Student since 2020
  • Filip Kovačević, PhD Student since 2024
  • Kevin Kögler, PhD Student since 2020
  • Peter Sukenik, PhD Student since 2021
  • Diyuan Wu, PhD Student since 2022
  • Clémentine Dominé, Postdoc since 2025

3 Research

We are at the center of a revolution in information technology, with data being the most valuable commodity. Exploiting this exploding number of data sets requires to address complex inference problems, and my research is aimed at developing mathematically principled solutions. These inference problems span different fields and arise in a variety of applications coming from engineering and natural sciences. In particular, in data science and machine learning, given a model for the observations, the focus is on how many samples convey sufficient information to perform a certain task and what are the optimal ways to utilize such samples. Both the vision and the toolkit that I adopt are inspired by information theory, and the long-term goal is to address the following fundamental questions: What is the minimal amount of information necessary to solve an assigned inference problem? Given this minimal amount of information, is it possible to design a low-complexity algorithm? How do the answers to these two questions change, if we add extra requirements in terms of robustness and privacy? To tackle these challenges, my work studies the dynamics of the gradient-based methods constituting the workhorse of deep learning, it characterizes the features being learnt, and it develops algorithms with information-theoretic optimality guarantees which are then applied to high-dimensional problems in quantitative genetics.


3.1 Current Scientific Projects/Directions

  • Theoretical foundations of trustworthy machine learning
  • Learning in high dimensions with differential privacy
  • Optimal inference across heterogeneous data
  • Information-theoretically principled algorithms for quantitative genetics

3.2 Research in Focus

In many deep learning applications, training datasets routinely include personal, sensitive information. Learning from these data is possible without creating privacy infringement via methods guaranteeing differential privacy, designed to provide provable protection to any individual user. However, differential privacy comes with a performance cost, and the cost is often believed to grow with the number of parameters of the learning model. Our work challenges this view, showing that over-parameterization is not at odds with privacy. In fact, we prove that, for a class of over-parameterized models having access to enough training samples, privacy even comes for free, i.e., with a small loss in performance. This result provides theoretical support to the development of differentially private models at scale. This work was published in the Proceedings of the National Academy of Sciences: S. Bombari and M. Mondelli, “Privacy for free in the overparameterized regime”, PNAS, vol. 122, no. 15, 2025.



theme sidebar-arrow-up
Back to Top