Local Learning Rules for Out-of-Equilibrium Physical Generative Models
This work addresses the challenge of training physical generative models for researchers in physics and machine learning, though it appears incremental as it applies known methods to new physical systems.
The paper tackled the problem of learning out-of-equilibrium driving protocols for score-based generative models using local learning rules, and demonstrated this by implementing a generative model in a network of driven oscillators to sample from Gaussian mixtures and generate MNIST digits 0 and 1.
We show that the out-of-equilibrium driving protocol of score-based generative models (SGMs) can be learned via local learning rules. The gradient with respect to the parameters of the driving protocol is computed directly from force measurements or from observed system dynamics. As a demonstration, we implement an SGM in a network of driven, nonlinear, overdamped oscillators coupled to a thermal bath. We first apply it to the problem of sampling from a mixture of two Gaussians in 2D. Finally, we train a 12x12 oscillator network on the MNIST dataset to generate images of handwritten digits 0 and 1.