A comment on the training of unsupervised neural networks for learning phases

The impact on the performance of an unsupervised neural network (NN) for learning the phases of two-dimensional ferromagnetic Potts model, namely a deep learning autoencoder (AE), from using various training sets is investigated. We find that data below and in the vicinity of the transition temperature T-c are crucial in training a successful AE. Our results also indicate that the commonly employed training procedures for unsupervised NNs are not efficient, and the obtained outcomes here can be considered as useful guidelines to set up effective trainings for unsupervised NNs.

Automated summary

This study investigates the impact of using various training sets on the performance of an unsupervised neural network (NN) for learning the phases of a two-dimensional ferromagnetic Potts model, specifically a deep learning autoencoder (AE). The results show that data below and near the transition temperature T-c are crucial in successfully training the AE. Additionally, the commonly used training procedures for unsupervised NNs are found to be inefficient, and the findings from this study can serve as useful guidelines for setting up effective trainings for unsupervised NNs.