Reproduce “Learning the Ising Model with Generative Neural Networks” on OpenResearch

About this paper

Francesco D'Angelo, Lucas Böttcher

This study quantitatively evaluates how Restricted Boltzmann Machines (RBMs) and Variational Autoencoders (VAEs) learn and reproduce the statistical distributions of the 2D Ising model across various temperatures. The research compares their ability to capture physical features like magnetization, energy, and spin-spin correlations, finding that while both perform well, RBMs show robustness near critical temperatures without explicit convolutions, and convolutional VAEs achieve better overall distribution similarity with multi-temperature training.

Key insights

• Convolutional layers are crucial for VAEs to effectively model spatial correlations in spin systems, allowing them to capture both coarse-grained and fine-grained features.
• RBMs demonstrate robust performance in generating samples and capturing physical features across all temperatures, including the challenging critical region, despite not using explicit convolutional layers.
• Training a single conditional VAE on data from all temperatures significantly improves its performance, especially near critical points, compared to training separate models for each temperature.

Summary by alphaXiv.