Connecting phase transition theory with unsupervised learning

TL;DR

This paper introduces an unsupervised learning approach that links phase transition theory with autoencoders, enabling automatic entropy estimation and order parameter identification, including complex spin glass phases.

Contribution

It demonstrates that autoencoder cross-entropy estimates entropy, proposes interpretable autoencoders for phase transition order parameters, and analyzes spin glass phases.

Findings

Autoencoder cross-entropy correlates with physical entropy.

Identified critical temperature via reconstruction loss inflection point.

Proposed a distributed order parameter for spin glasses.

Abstract

Entropy and order parameter are two key concepts in phase transition theory. This paper proposes an unified method to both find order parameter and estimate entropy automatically with unsupervised learning. The contributions of this paper are threefold: First, it is shown that the cross-entropy loss of an optimum autoencoder could be used to estimate the physical entropy, which also explains why the critical temperature can be identified by the inflection point of the reconstruction loss. Second, a series of interpretable autoencoders are proposed which reproduce the ferromagnetic/antiferromagnetic (F/AF) order parameter in special cases. They provide us an intuitive prototype to understand the connection between unsupervised learning and phase transition theory. Third, we analyze spin glass phase with our method, the results suggest a "distributed" order parameter to describe spin-glass ground state, which is a natural generalization of F/AF order parameter with respect to the autoencoder.