Font Size:
Can Neural-Networks learn a Gauge symmetry ?
##manager.scheduler.building##: Edificio San Jose
##manager.scheduler.room##: Aula Juan Pablo II
Date: 2019-07-11 05:00 PM – 05:15 PM
Last modified: 2019-06-09
Abstract
During the past decades, the field of Machine Learning has been very successful in designing mathematical models and learning algorithms to learn
automatically how to model data in order to obtain latent representation leading to incredible performance both in classification tasks and in gener-
ating credible new data from a data-set. Such powerful tools have potentially a broad application in many other fields, and in particular in Statistical
Physics. In this work, we explore the possibility of using neural networks to predict the thermodynamic properties of the spin glass Edwards-Anderson
model, just by looking at the interaction network. If such a machine existed, it should be able to detect a complex non-local symmetry that couples the
physical properties of large groups of disorder realizations (samples): a gauge symmetry. We tackle the problem of designing a neural
network able to distinguish if two samples are, or not, related by a gauge transformation.
automatically how to model data in order to obtain latent representation leading to incredible performance both in classification tasks and in gener-
ating credible new data from a data-set. Such powerful tools have potentially a broad application in many other fields, and in particular in Statistical
Physics. In this work, we explore the possibility of using neural networks to predict the thermodynamic properties of the spin glass Edwards-Anderson
model, just by looking at the interaction network. If such a machine existed, it should be able to detect a complex non-local symmetry that couples the
physical properties of large groups of disorder realizations (samples): a gauge symmetry. We tackle the problem of designing a neural
network able to distinguish if two samples are, or not, related by a gauge transformation.