Portrait of Francesco Tonin

Francesco Tonin

Machine Learning & AI

Email: francesco.tonin@epfl.ch

I am a postdoctoral fellow at EPFL as part of the LIONS lab, hosted by Prof. Volkan Cevher.

I obtained my PhD in Machine Learning at KU Leuven in ESAT-STADIUS, advised by Prof. Johan Suykens and Prof. Panos Patrinos. Previously, I received my MSc in Computer Science from KU Leuven (2019) and my BSc in Computer Engineering from Politecnico di Torino (2017).

My research focuses on developing algorithms for data- and compute-efficient machine learning, with applications to language and multi-modal models. My work spans:

  • Data selection & mixing: Understand which training data is most useful for building better models, and how to combine data from different sources (Chameleon) and types, such as text, images, and video (MaD-Mix). A broader introduction is available in my seminar on data mixing.

  • Optimization: Develop algorithms that make model pre-training faster and more effective (AcceleGrad#), and explore mathematical reformulations that enable efficient optimization under additional constraints such as sparsity and robustness (kernel PCA through dualization) and fairness (fair spectral clustering).

  • Model architecture: Explore scalable ways to reduce inference costs (Neural Block Linearization), enable highly parameter-efficient adaptation (Quantum-PEFT), and develop alternative attention mechanisms based on bidirectional linear Transformers (LION) and asymmetric kernel SVD (Primal-Attention).

My longer-term goal is to understand how data, optimization, and architecture can be co-designed to improve the training of large-scale AI models. See my research overview for more detail and my updated CV. You can find me on GitHub and LinkedIn.

Selected publications

For the full list, see all publications.

News

  1. Awarded 20k GPU hours (GH200) from Swiss AI to study quantization-aware training.

  2. I will be serving as Area Chair for ICLR 2027.

  3. Two review papers accepted on deep learning architectures and kernel SVD.

  4. Recognized as Top Reviewer at ICML 2026.

  5. MaD-Mix accepted as oral presentation at DATA-FM @ ICLR 2026.

  6. Gave a seminar on data mixing in LLMs at ESAT-STADIUS, KU Leuven.

  7. Awarded a 50k GPU-hour (GH200) compute grant from Swiss AI to investigate online data pipelines in LLMs.

  8. Recognized as Top Reviewer at NeurIPS 2025.