Emergence of Nonequilibrium Latent Cycles in Unsupervised Generative Modeling

Abstract

We show that nonequilibrium dynamics can play a constructive role in unsupervised machine learning by inducing the spontaneous emergence of latent-state cycles. We introduce a model in which visible and hidden variables interact through two independently parametrized transition matrices, defining a Markov chain whose steady state is intrinsically out of equilibrium. Likelihood maximization drives this system toward nonequilibrium steady states with finite entropy production, reduced self-transition probabilities, and persistent probability currents in the latent space. These cycles are not imposed by the architecture but arise from training, and models that develop them reproduce the empirical distribution of data classes more faithfully, with a clear correlation between agreement with the data and entropy production. Compared with equilibrium approaches such as restricted Boltzmann machines, our model breaks the detailed balance between the forward and backward conditional transitions and relies on a log-likelihood gradient that depends explicitly on the last two steps of the Markov chain. Hence, this exploration of the interface between nonequilibrium statistical physics and modern machine learning suggests that introducing irreversibility into latent-variable models can improve the fidelity of the generated data distribution.

v3: 11 pages, 6 figures. Accepted in PRE. Revised example with respect to v2

References in corpus (19)

• A high-bias, low-variance introduction to Machine Learning for physicists
• Non-equilibrium statistical mechanics: From a paradigmatic model to biological transport
• Phase Coexistence in Driven One Dimensional Transport
• Event-chain Monte Carlo algorithms for hard-sphere systems
• Efficiency of cellular information processing
• Generalized event-chain Monte Carlo: Constructing rejection-free global-balance algorithms from infinitesimal steps
• Emergence of Compositional Representations in Restricted Boltzmann Machines
• Thermodynamics of Error Correction
• Phase transitions in Restricted Boltzmann Machines with generic priors
• Dynamical Regimes of Diffusion Models
• Computational protein design with evolutionary-based and physics-inspired modeling: current and future synergies
• Equilibrium and non-Equilibrium regimes in the learning of Restricted Boltzmann Machines
• Nonequilibrium Thermodynamics of Restricted Boltzmann Machines
• Unsupervised hierarchical clustering using the learning dynamics of RBMs
• Barriers and Dynamical Paths in Alternating Gibbs Sampling of Restricted Boltzmann Machines
• Kinetics and thermodynamics of DNA polymerases with exonuclease proofreading
• Interpretable machine learning of amino acid patterns in proteins: a statistical ensemble approach
• Unlearning regularization for Boltzmann Machines
• Speed-accuracy relations for diffusion models: Wisdom from nonequilibrium thermodynamics and optimal transport