Transfer Learning for Larger, Broader, and Deeper Neural-Network Quantum States

Remmy A. M. Zen, Stéphane Bressan

Abstract: Neural-network quantum states are a family of unsupervised neural network models simulating quantum many-body systems. We investigate the efficiency and effectiveness of neural-network quantum states with deep restricted Boltzmann machine with different sizes, breadths, and depths. We propose and evaluate several transfer learning protocols for the improvement of scalability, effectiveness, and efficiency of neural-network quantum states with different numbers of visible nodes, hidden nodes per layer, and hidden layers. The results of a comparative empirical performance evaluation confirm the advantages of deep neural-network quantum states and of the proposed transfer learning protocols.