Journées Lensoise de Géométrie et Topologie

Programme

Lundi :

12:30 Déjeuner à l'atelier Marc Meurin

14:00 Mihaly Petreczky

15:00 Benoit Dherin (visio-conférence)

Mardi :

09:30 François Petit

10:30 Gennaro di Brino (visio-conférence)

11:30 Mathieu Klimczak

12:30 Déjeuner à l'Ardoise (formulaire)

14:00 Nicolas Ricka

15:00 Thibault Defourneau

All talks (given in person or remotely) will be on zoom at the link:

https://univ-artois-fr.zoom.us/j/92303799154

Topic: Journees Lensoises

Time: Jul 12, 2021 01:45 PM Paris

Meeting ID: 923 0379 9154

Les exposés seront donné dans l'amphi S25. Pour y accéder, à partir de l'entrée principale du bâtiment B, descendre au sous-sol, ce sera l'amphi le plus à droite.

Titres et résumés

Mathieu Klimczak :

I did a PhD in algebraic topology, and then all of a sudden I was advertised as an "expert in Artificial Intelligence". The idea of this talk will be to briefly present the company I work at (CITC-EuraRFID), and the changes I noticed by switching from academia to the private sector. I will then use most of my time to talk about 2 uses cases on embedded computer vision I had to work on.

Thibault Defourneau :

The main idea of this talk is to share my experience from Mathematics to Data science, and to explain the challenges I faced in the digital word. First, I will present briefly Trinov, the company I am working for, as well as main solutions developed by exploiting data. Then, I will focus on one specific solution I develop by explaining some key concepts and techniques behind it.

Benoit Dherin

Title: Implicit Gradient Regularization

Abstract: Large deep neural networks used in modern supervised learning have a large submanifold of interpolating solutions, most of which are not good. However, it has been observed experimentally that gradient descent tends to converge in the vicinity of flat interpolating solutions producing trained models that generalize well to new data points, and the more so as the learning rate increases. Using backward error analysis, we will show that gradient descent actually follows the exact gradient flow of a modified loss surface, which can be described by a regularized loss preferring optimization paths with shallow slopes, and in which the learning rate plays the role of a regularization rate. (This is joint work with David Barrett from DeepMind).

Gennaro Di Brino

Title: An incomplete survey of recommender systems

Abstract: We will give a high level overview of the kind of models that do the heavy lifting in most of the major tech companies, by suggesting items, content, movies, etc.. to users. In doing so, we will mention some interesting techniques that transformed the field of Data Science, and some other very promising ones. The focus will be on the shift of perspective it takes to view yet the same mathematical object or even the same model architecture in a completely different context, where it can be extremely effective.

Mihaly Petreczky (CNRS, Ecole Centrale Lille, research group CIRStAL).

Title: Control theory and machine learning for dynamical systems.

Abstract: In this talk I will present the relationship between control theory and learning dynamical systems, and I will present some mathematical problems

which arise in this setting. The goal of control theory is to develop algorithms for influencing the behavior of dynamical systems. To this end,

models of dynamical systems are required. In turn models are often derived from measurement data.

The branch of control theory which studies the estimation of

models from dynamical systems using data is known as system identification. The goal of machine learning is to build models from data. In particular,

many popular model classes (recurrent neural networks, LSTM) used in machine learning are in fact models of dynamical systems.

Hence, the corresponding theoretical challenges are the same as in system identification. In particular, one of the theoretical tools to prove

properties of learning algorithms is realization theory, which addresses the relationship between the input-output behavior of the model and its

internal structure. In turn, realization theory gives rise to several interesting mathematical problems which are interesting on their own right.

Nicolas Ricka

Title: Topological analysis for mental health prediction

Abstract: In this talk, we will motivate the use of machine learning to study certain aspects of mental health. In the specific case of major depressive disorder and post traumatic stress disorder, we will discuss the advantages of using topological invariants to help the clinician in assessing the patients, and describe some results in this direction.

Participant.e.s

Ivo Dell'Ambrogio (Lille) m

Thierry Bay (UPHF) m

Ouriel Bloede (Angers)m

Benoit Brebion lm

Thibault Defourneau (Trinov) l

Benoit Dhérin (Google)

Gennaro Di Brino (Come)

Naim Es Sebani

Yaël Frégier (Artois) lm

Matthieu Goliot lm

Martin Gonzalez

Jean-Baptiste Gouray (Artois) lm

Arthur Kipfel (Artois) lm

Mathieu Klimcak (Eurafid) lm

Anthony Pereira lm

Olivier Peltre (Artois)

Mihaly Petreczky (CNRS) l

Camilla Penzo (Artois) lm

François Petit (Paris) m

Germain Poloudenny l

Nicolas Ricka p

Martin Saralegi (Artois) lm

Arthur Soulié (Glasgow)

Daniel Tanré (Lille)

Informations pratiques

LML - Laboratoire de Mathématiques de Lens
http://lml.univ-artois.fr/public/francais/index.html
Faculté des Sciences Jean Perrin, Université d'Artois
Rue Jean Souvraz SP 18, 62307 LENS CEDEX

Nous avons des fonds pour financer partiellement votre séjour

Les organisateurs

Ivo Dell'Ambroggio, Yaël Fregier, Camilla Penzo, Martin Saralegi

Inscription

Contacter camilla.penzo@univ-artois.fr

Remerciements

Nous remercions la fédération de mathématiques du Nord ainsi que le laboratoire de mathématiques de Lens pour leur soutien financier.