Pantheon SEMPARIS Le serveur des séminaires parisiens Paris

Status Confirmed
Seminar Series IPHT-STA
Subjects cond-mat
Date Monday 2 October 2017
Time 11:00
Institute IPHT
Seminar Room Salle Claude Itzykson, Bât. 774
Speaker's Last Name Aurélien Decelle
Speaker's First Name
Speaker's Email Address
Speaker's Institution Lab. de Recherche en Informatique, Univ. Paris Sud, Orsay
Title Spectral learning of Restricted Boltzmann Machines
Abstract In this presentation I will expose our recent results on the Restricted Boltzman Machine (RBM). The RBM is a generative model very similar to the Ising model, it is composed of both visible and hidden binary variables, and traditionally used in the context of machine learning. In this context, the goal is to inferred the parameters of the RBM such that it reproduces correctly a dataset's distribution. Although they have been widely used in computer science, the phase diagram of this model is not known precisely in the context of learning. In particular, it is not known how the parameters influence the learning, and what exactly is learned within the parameters of the model. In our work, we show how the SVD of the data governs the first phase of the learning and how this decomposition helps to dynamics and the equilibrium properties of the model.
arXiv Preprint Number
Comments
Attachments

To Generate a poster for this seminar : [ Postscript | PDF ]

[ Bulletin Board ]    [ Subscription ]    [ Archive]    [ Help ]    [ ] [ Version francaise ]