LABORATOIRE DE PHYSIQUE THEORIQUE DE LA MATIERE CONDENSEE

 

 

Attention : désormais les séminaires ont lieu tous les lundis à 10h45 en salle  523 du LPTMC - Tour 12-13 


 

Marylou Gabrié (LPS-ENS Paris)

Entropy and mutual information in models of deep neural networks


The successes and the multitude of applications of deep learning methods have spurred efforts towards quantitative modeling of the performance of deep neural networks. In particular, an information-theoretic approach has been receiving increasing interest. Nevertheless, it is in practice computationally intractable to compute entropies and mutual informations in industry-sized neural networks. In this talk, we will consider instead a class of models of deep neural networks, for which an expression for these information-theoretic quantities can be derived from the replica method. We will examine how mutual informations between hidden and input variables can be reported along the training of such neural networks on synthetic datasets. Finally we will discuss the numerical results of a few training experiments.

ATTENTION SALLE INHABITUELLE: bibliothèque du LPTHE (4ème étage tour 13, couloir 13-14):

Yshai Avishai (Ben Gurion University, Beer Sheva, Israel)

Phase Aharonov-Casher dans les systèmes mésoscopiques

La phase Aharonov-Casher joue un rôle important dans les systèmes mésoscopiques dans lesquels le couplage spin-orbite est pertinent. Dans cette présentation, nous examinons la dépendance des observables physiques pertinentes sur la phase Aharonov-Casher (telles que la conductance g et la polarisation électronique P). Premièrement, nous suggérons une expression de la phase qui est manifestement invariante de jauge. Ensuite, nous considérons un problème de diffusion dans lequel la phase dépend d’au moins deux paramètres x et y. Notre résultat principal est que la conductance g dépend des deux paramètres uniquement à travers la phase Aharonov-Casher, alors que la polarisation P dépend de chaque paramètre séparément.

ATTENTION SALLE INHABITUELLE: salle 3-17, couloir 23-22, 3ème étage (grande salle INSP)

Deux docteurs dans le monde de l'entreprise

Boris Mantisi (Quatorze-IG, Paris) et Simon Moulieras (MyndBlue, Polytechnique)

La formation universitaire reste méconnue de la majorité des acteurs du secteur privé et les liens entre ces deux mondes sont encore trop peu développés. Pourtant, la formation universitaire, et notamment au travers de la réalisation d'une thèse, offre de nombreux atouts que le docteur doit absolument apprendre à valoriser afin de s'ouvrir les portes du privé. Deux docteurs en physique, anciens post-docs du LPTMC, interviennent pour partager leur saut vers l'entreprise. Simon Moulieras, data scientist chez MyndBlue, à Polytechnique et Boris Mantisi, data scientist chez Quatorze IG, bureau d'études d'ingénierie basé dans le 13 ème arrondissement de Paris. Chacun parlera 20 minutes et il y aura ensuite du temps de discussion.

 

Andrea De Luca (Oxford)

Solution of a minimal model for many-body quantum chaos

I will present a minimal model for quantum chaos in a spatially extended many-body system. It consists of a chain of sites with nearest-neighbour coupling under Floquet time evolution. Quantum states at each site span a q-dimensional Hilbert space and the time evolution is specified as a random circuit, which is random in space but periodic in time (Floquet). Each site is coupled via a random matrix to its neighbour on one side during the first half of the evolution period, and to its neighbour on the other side during the second half of the period. I will introduce a diagrammatic formalism useful to average the many-body dynamics over realisations of the random matrices. This approach leads to exact expressions in the large-q limit and sheds light on the universality of random matrices in many-body quantum systems and the ubiquitous entanglement growth in out-of-equilibrium dynamics. I will also discuss universal behaviour which goes beyond random matrix theory and the role played by space dimensionality which emerges through a mapping into the classical Potts model, exact at large q.

Thierry Dauxois (ENS Lyon & CNRS)

Energy cascade in internal wave attractors

Internal gravity waves play a primary role in geophysical fluids : they contribute significantly to mixing in the ocean and they redistribute energy and momentum in the middle atmosphere. In addition to their very interesting and very unusual theoretical properties, these waves are linked to one of the important questions in the dynamics of the oceans: the cascade of mechanical energy in the abyss and its contribution to mixing.

Combining the physics of waves, dynamical systems theory and oceanography, I will discuss a unique self-consistent experimental and numerical setup that models a cascade of triadic interactions transferring energy from large-scale monochromatic input to multi-scale internal wave motion. I will also provide explicit evidence of a wave turbulence framework for internal waves. Finally, I will show how beyond this regime, we have a clear transition to a cascade of small-scale overturning events which induce mixing.