Efficient training of energy-based models via spin-glass control

TL;DR

This paper proposes a new family of energy-based models inspired by spin-glass control, enabling faster and more efficient training of Boltzmann machines for unsupervised learning tasks like MNIST.

Contribution

Introducing a novel energy-based model family based on spin-glass properties that simplifies training and improves efficiency over traditional methods.

Findings

Achieves comparable performance to standard methods on Bars and Stripes and MNIST datasets.

Eliminates the need for Markov chain Monte Carlo sampling in training.

Facilitates easy access to low-energy configurations for efficient learning.

Abstract

We introduce a new family of energy-based probabilistic graphical models for efficient unsupervised learning. Its definition is motivated by the control of the spin-glass properties of the Ising model described by the weights of Boltzmann machines. We use it to learn the Bars and Stripes dataset of various sizes and the MNIST dataset, and show how they quickly achieve the performance offered by standard methods for unsupervised learning. Our results indicate that the standard initialization of Boltzmann machines with random weights equivalent to spin-glass models is an unnecessary bottleneck in the process of training. Furthermore, this new family allows for very easy access to low-energy configurations, which points to new, efficient training algorithms. The simplest variant of such algorithms approximates the negative phase of the log-likelihood gradient with no Markov chain Monte Carlo sampling costs at all, and with an accuracy sufficient to achieve good learning and generalization.