Maria Refinetti
PhD @ ENS Paris
I am PhD student at Ecole Normale Supérieure in Paris under the supervision of Florent Krzakala. My research focuses on applying tools from theoretical physics to understand the success of modern machine learning.
Previously, I earned a Bachelor and Master degrees in Physics at Ecole Polytechnique Fédérale de Lausanne. My main focus was on theoretical high energy physics. My master thesis, led by Alessandro Vicchi, explored conformal bootstrap methods.
Research Interests
Machine Learning is very successful in many aspects of daily life. My PhD, in statistical mechanics of learning aims to give some theoretical understanding of its success.
In one line of research, I work towards understanding the generalisation properties of over-parametrised neural network. For example, one of my works aims to characterise the role of the various sources of noise in the decrease of the test error in the over-parametrised regime. Another work, compares the performances of lazy methods such as kernel methods to those of neural networks on synthetic datasets.
I am also interested in describing the dynamics of Neural Networks. This allows to study alternative training algorithm, such as direct feedback alignment which tries to circumvent some of the pitfall of vanilla back propagation and is more biologically plausible. I am also working towards applying these dynamical methods to study toy models of unsupervised learning such as non-linear auto-encoders.
If you want to know more scroll down to some view some talks, read my publications or email me at mariaref@gmail.com .
S. d’Ascoli, M. Refinetti, G. Biroli, & F. Krzakala
International Conference on Machine Learning, PMLR, 2020 [Slides]
M. Refinetti, S. Goldt, F. Krzakala & L. Zdeborová
International Conference on Machine Learning, PMLR, 2021 [Poster] [Slides]
M. Refinetti, S. d’Ascoli, R. Ohana, & S. Goldt
International Conference on Machine Learning, PMLR, 2021 [Poster] [Slides]
Proceedings of the National Academy of Sciences 118.32 (2021)
Talks
DeepMath 2021
Check out 5:27 to see my Talk on The dynamics of learning in shallow non linear Auto-encoders at DeepMath 2021.
Les Houches Workshop 2020
Go to 18:00 to listen to a talk on the Bias Variance trade-off in modern machine learning [Slides]
Education [CV]
PhD at Ecole Normale Supérieure Paris
PhD under the supervision of Prof. F. Krzakala on the study of Artificial Neural Networks. Research combining analytical methods to practical implementation.
The Learning Dynamics of Neural Networks and Optimization problems: a Spin Glass Approach
Master & Bachelor at Ecole Polytechnique Fédérale de Lausanne
Final grade 5.6/6
Bachelor in Physics: Final grade 5.3/6
Graduated in the best 5% of the class
Scuola Europea di Varese
European Baccalaureate: Final Grade 92/100
Scientific Orientation