Filip KovačevićI am a PhD student at ISTA, supervised by Prof. Marco Mondelli. Before that, I have obtained a MS in Mathematics from ETH Zurich, and a BS in Mathematics from University of Belgrade. I am interested in the theoretical foundations of machine learning, using tools and ideas from high-dimensional statistics, random matrix theory, and learning theory. Recently, I’ve been focusing broadly on questions of sample efficiency, and particularly data selection, transfer learning, and the role of data reuse in optimization. Email / CV / Google Scholar / LinkedIn |
|
Publications and Preprints |
High-dimensional Asymptotics and Dataset Selection for Private Transfer LearningFilip Kovačević, Edwige Cyffers, Stefano Sarao Mannelli, Marco Mondelli preprint, 2026 arxiv / We give a precise analysis of the test error under private transfer learning through the lens of high-dimensional regression, allowing for beneficial dataset selection using only summary statistics. |
Full-Batch Gradient Descent Outperforms One-Pass SGD: Sample Complexity Separation in Single-Index LearningFilip Kovačević, Hong Chang Ji, Denny Wu, Mahdi Soltanolkotabi, Marco Mondelli ICML, 2026 arxiv / video / poster / For a quadratic single-index model, we identify when full-batch gradient descent achieves recovery with lower sample complexity than one-pass SGD. |
Towards a holistic understanding of Selection Bias for Causal Effect IdentificationYiwen Qiu, Filip Kovačević, Shimeng Huang, Peter Spirtes, Francesco Locatello ICML, 2026 arxiv / video / We provide necessary and sufficient condition for ATE identifiability under selection bias. |
High-dimensional Analysis of Synthetic Data SelectionParham Rezaei, Filip Kovačević, Francesco Locatello, Marco Mondelli ICLR, Oral, 2026 arxiv / video / poster / We analyze data augmentation with generative models from a random matrix theory perspective, revealing the importance of matching the covariance when augmenting real data. |
Spectral Estimators for Multi-Index Models: Precise Asymptotics and Optimal Weak RecoveryFilip Kovačević, Yihan Zhang, Marco Mondelli COLT, 2025 arxiv / video / poster / We provide a precise analysis of spectral methods for recovering low-dimensional signal subspaces in multi-index models, revealing a phase transition phenomenon that determines exactly when these methods succeed and enabling minimization of the needed sample size. |
Learning Pareto manifolds in high dimensions: How can regularization help?Tobias Wegel, Filip Kovačević, Alexandru Ţifrea, Fanny Yang AISTATS, 2025 arxiv / poster / We discuss how the application of vanilla regularization approaches can fail, and propose a two-stage MOL framework that can successfully leverage low-dimensional structure. |
Teaching and ServiceReviewer (Conferences): ICML 2026 (Gold Reviewer), ICLR 2026 Reviewer (Journals): TMLR, Mathematical Foundations of Machine Learning ETH Zurich, TA: Linear Algebra (Spring 2022), Algebra 1 (Spring 2023), Algebra 2 (Fall 2023) University of Belgrade, TA: Linear Algebra (Fall 2020) |
|
Design and source code from Leonid Keselman's website |