Partenaires

MSC
Logo CNRS Logo Université Paris Diderot
Logo tutelle Logo tutelle



Rechercher

Sur ce site

Sur le Web du CNRS


Accueil du site > Seminars > Séminaires théorie > Theory Club Monday March 9 2020 at 12:00 in room 734A. Takahiro Nemoto: "Optimizing inter-particle interactions using large deviation theory: from a viewpoint of Bayesian inference with regularization".

Theory Club Monday March 9 2020 at 12:00 in room 734A. Takahiro Nemoto: "Optimizing inter-particle interactions using large deviation theory: from a viewpoint of Bayesian inference with regularization"

Unless otherwise stated, seminars and defences take place at 11:30 in room 454A of Condorcet building.


Optimizing inter-particle interactions using large deviation theory: from a viewpoint of Bayesian inference with regularization

Takahiro Nemoto

Abstract: Bayesian inference plays a prominent role in many fields outside physics, such as quantitative political science, epidemiology and bioinformatics, where input data are typically complex, obtained in field researches or experiments. When modeling only essential properties of dynamics, as often important in statistical physics, input data are greatly reduced to just a few variables. In such cases, the Bayesian approach results in non-unique and inconsistent outcomes that are sensitive to the prior. In this talk, I will show a regularization technique based on the large deviation theory, in order to construct a set of priors to get a unique outcome from a single time-averaged input.

We will then apply this Bayesian inference to active Brownian particles with repulsive interactions. In our Bayesian framework, the only input we specify is to reduce the collisions among the particles. The framework then automatically selects (additional) inter-particle interactions to achieve this input in an optimal manner. With the help of the population dynamics algorithm, which is a technique used in the large deviation theory, we will perform this inference. We will find that the obtained optimal inter-particle interactions produce collective motions, even if the original particles do not have any aligning interactions. In other words, the collective motion will be obtained without explicitly modeling the details of inter-particle interactions, but only with a simple and practical input: reducing the collisions among particles within our Bayesian framework.

At the end of this talk, I will describe future perspectives of application of this framework in biology and engineering.

Monday March 9 at 12:00 in room 734A


Contact : Équipe séminaires / Seminar team - Published on / Publié le 6 March 2020


In the same section :