The unsupervised machine learning is generally useful to extract features
of input data. Since it can be regarded as a kind of information compression,
some researchers suggest its similarity to coarse-graining and renormalization.
In this talk, we use the spin configurations of Ising model as the input data
and the restricted Boltzmann machine (RBM) as the method of unsupervised
learning. Then we look at what kind of features the machine extracts, using
our method of “RBM flow”.
As a result, we can find an interesting similarity to renormalization and some
coincidence with thermodynamics. However, we also discover apparent
differences from renormalization and then argue why such phenomena occur
by studying the parameter dependence in machine learning.