Learning Stochastic Representations of Physical Systems

Abstract

Learning representations of physical systems is an important problem at the interface of statistical physics and machine learning. Recently, there has been a growing interest in devising methods to analyze high-dimensional simulation data generated by unbiased or biased samplers. As statistical physics systems consisting of N 1 objects tend to have many degrees of freedom, dimensionality reduction methods are of particular interest. Here, we use a new method, multiscale reweighted stochastic embedding (MRSE), to analyze handwritten digits data sets and a biased trajectory of alanine tetrapeptide, and show that we can reconstruct low-dimensional representations of these data sets while retaining the most informative characteristics of their high-dimensional representation.

Tasks

Reproductions