Indaco Biazzo — EPFL #
Boltzmann Autoregressive Neural Networks #
Generative Autoregressive Neural Networks (ARNNs) excel in generative tasks across various domains, including images, language, and science. Particularly in physics, they have successfully applied to generate samples from statistical physics models.  Despite their success, ARNN architectures often operate as black boxes without a clear connection to underlying physics or statistical models. This seminar explores the direct link between neural network architectures and physics models. I'll show how the neural network parameters align with Hamiltonian couplings and external fields, highlighting the emergence of residual connections and recurrent architectures from the derivation. By leveraging statistical physics techniques, we formulate ARNNs for specific systems, and I’ll discuss a new approach for sampling from sparse interacting systems, crucial for physics, optimization, and inference problems. Our findings validate a physically informed approach and suggest potential extensions to multivalued variables, paving the way for broader applications in scientific research.