معرفی
Raphaël BERTHIER is a researcher affiliated with INRIA and Sorbonne Université, actively contributing to theoretical machine learning research through seminars at institutions like CREST-CMAP.
His work centers on the implicit regularization properties of diagonal linear networks, demonstrating their mathematical equivalence to the lasso regularization path where training time functions as an inverse regularization parameter. This research bridges neural network optimization with sparse modeling theory, revealing how early stopping induces sparsity in linear models without explicit penalization.
As an INRIA researcher, he operates within France's premier digital science research ecosystem, focusing on rigorous theoretical foundations for deep learning dynamics while collaborating with academic centers like Sorbonne Université's mathematics and computer science divisions.
