A universal training strategy for machine learning phases and criticalities

Yuan-Heng Tseng^1*, Fu-Jiun Jiang¹, Ching-Yu Huang²

¹Department of Physics, National Taiwan Normal University, Taipei, Taiwan
²Department of Applied Physics, Tunghai University, Taichung, Taiwan

* Presenter:Yuan-Heng Tseng, email:a55664389@gmail.com

In order to find the critical point of phase transition, one needs to be able to distinguish different phases. In machine learning, this can be viewed as a classification problem. In this talk, I will introduce a universal training strategy to train unsupervised neural networks: an autoencoder (AE) and a generative adversarial network (GAN) are trained only once on a one-dimensional lattice of 200 sites. Two artificially made configurations are used as the training set. In addition, the AE contains only one hidden layer with two neurons and both the generator and the discriminator of the GAN have two neurons as well. Although the structure of neural networks is extremely simple, it can determine the critical points of various models precisely. These models include the two-dimensional two-state Potts model, the two-dimensional generalized classical XY model, and the three-dimensional classical Ο(3) model.

Keywords: phase transition, neural network, classical spin system