Publications by Maria Refinetti

Neural networks trained with SGD learn distributions of increasing complexity

The uncanny ability of over-parameterised neural networks to generalise well has been explained using various ‘simplicity biases’. These theories postulate that neural networks avoid overfitting by first fitting simple, linear classifiers before learning m ...

2025

Bootstrapping traceless symmetric O(N) scalars

Alessandro Vichi, Maria Refinetti, Marten Jan Reehorst

We use numerical bootstrap techniques to study correlation functions of traceless sym-metric tensors of O(N) with two indices ti j. We obtain upper bounds on operator dimen-sions for all the relevant representations and several values of N. We discover sev ...

SCIPOST FOUNDATION2023

Align, then memorise: the dynamics of learning with feedback alignment*

Maria Refinetti

Direct feedback alignment (DFA) is emerging as an efficient and biologically plausible alternative to backpropagation for training deep neural networks. Despite relying on random feedback weights for the backward pass, DFA successfully trains state-of-the- ...

2022

The dynamics of representation learning in shallow, non-linear autoencoders

Maria Refinetti

Autoencoders are the simplest neural network for unsupervised learning, and thus an ideal framework for studying feature learning. While a detailed understanding of the dynamics of linear autoencoders has recently been obtained, the study of non-linear aut ...

JMLR-JOURNAL MACHINE LEARNING RESEARCH2022

Epidemic mitigation by statistical inference from contact tracing data

Florent Gérard Krzakala, Lenka Zdeborová, Maria Refinetti, Indaco Biazzo, Giovanni Catania

Contact tracing is an essential tool to mitigate the impact of a pandemic, such as the COVID-19 pandemic. In order to achieve efficient and scalable contact tracing in real time, digital devices can play an important role. While a lot of attention has been ...

National Academy of Sciences2021

Classifying high-dimensional Gaussian mixtures: Where kernel methods fail and neural networks succeed

Florent Gérard Krzakala, Lenka Zdeborová, Maria Refinetti

A recent series of theoretical works showed that the dynamics of neural networks with a certain initialisation are well-captured by kernel methods. Concurrent empirical work demonstrated that kernel methods can come close to the performance of neural netwo ...

PMLR2021

Align, then memorise: the dynamics of learning with feedback alignment

Maria Refinetti

Direct Feedback Alignment (DFA) is emerging as an efficient and biologically plausible alternative to backpropagation for training deep neural networks. Despite relying on random feedback weights for the backward pass, DFA successfully trains state-of-the- ...

JMLR-JOURNAL MACHINE LEARNING RESEARCH2021