Efficient training of energy-based models via spin-glass control
This addresses the training bottleneck for unsupervised learning models, offering potential for more efficient algorithms, though it appears incremental in scope.
The paper tackles the inefficient training of energy-based models like Boltzmann machines by introducing a new family of models that control spin-glass properties, showing they quickly achieve performance comparable to standard methods on datasets like Bars and Stripes and MNIST.
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.