publications
* denotes co-first authorship. Check also my Google Scholar profile.
2025
-
The physics of data and tasks: Theories of locality and compositionality in deep learningPhD Dissertation, École Polytechnique Fédérale de Lausanne (EPFL), 2025 -
Bigger isn’t always memorizing: Early stopping overparameterized diffusion modelsarXiv preprint, 2025 -
Scaling laws and representation learning in simple hierarchical languages: Transformers vs. convolutional architecturesarXiv preprint, 2025 -
MEMOIR: Lifelong model editing with minimal overwrite and informed retention for LLMsAdvances in Neural Information Processing Systems (NeurIPS), 2025 -
Backdoor unlearning through linear task decompositionICML 2025 Workshop on Machine Unlearning for Generative AI, 2025 -
How compositional generalization and creativity improve as diffusion models are trainedInternational Conference on Machine Learning (ICML), PMLR, 2025 -
Probing the latent hierarchical structure of data via diffusion modelsInternational Conference on Learning Representations (ICLR), 2025 -
LiNeS: Post-training layer scaling prevents forgetting and enhances model mergingInternational Conference on Learning Representations (ICLR), 2025 -
A phase transition in diffusion models reveals the hierarchical nature of dataProceedings of the National Academy of Sciences (PNAS), 2025
2024
-
Multi-modal hallucination control by visual information groundingIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024 -
Task addition and weight disentanglement in closed-vocabulary modelsICML 2024 Efficient Systems for Foundation Models Workshop, 2024 -
How deep neural networks learn compositional data: The Random Hierarchy ModelPhysical Review X, 2024 -
Computational complexity of deep learning: fundamental limitations and empirical phenomenaJournal of Statistical Mechanics: Theory and Experiment (JSTAT), 2024
2023
-
What can be learnt with wide convolutional neural networks?International Conference on Machine Learning (ICML), PMLR, 2023
2021
-
Locality defeats the curse of dimensionality in convolutional teacher-student scenariosAdvances in Neural Information Processing Systems (NeurIPS), 2021 -
Relative stability toward diffeomorphisms indicates performance in deep netsAdvances in Neural Information Processing Systems (NeurIPS), 2021
2020
-
Spectral analysis of infinitely wide convolutional neural networksMaster’s Thesis, Sorbonne Université and Politecnico di Torino, 2020
