Unsupervised machine learning for physical concepts

TL;DR

This paper introduces a hybrid unsupervised machine learning approach to extract interpretable physical concepts from experimental data by combining topological analysis with neural networks.

Contribution

It presents a novel two-stage method that uses Betti numbers and variational autoencoders to identify physical concepts without supervision.

Findings

Successfully applied to toy models demonstrating concept extraction.

Shows potential for aiding scientific discovery through unsupervised learning.

Provides a framework for interpreting experimental data in physical sciences.

Abstract

In recent years, machine learning methods have been used to assist scientists in scientific research. Human scientific theories are based on a series of concepts. How machine learns the concepts from experimental data will be an important first step. We propose a hybrid method to extract interpretable physical concepts through unsupervised machine learning. This method consists of two stages. At first, we need to find the Betti numbers of experimental data. Secondly, given the Betti numbers, we use a variational autoencoder network to extract meaningful physical variables. We test our protocol on toy models and show how it works.