Annealed Gibbs Measures with Data-Driven Interaction Potentials on Hierarchical Structures
L. U. Abdullaev, F. Herrera, U. A. Rozikov, M. V. Velasco
原文摘要
We construct annealed Gibbs measures on continuous interaction parameters and binary spins on Cayley trees of order $k\geq 1$, with coupling functions determined by an empirical loss. The interaction parameters and spins are distinct random variables: the former determine edge couplings, while the latter represent binary labels. The training data specify the coupling function and the joint Gibbs specification, rather than automatically selecting a unique Gibbs measure on interaction parameters. We derive necessary and sufficient compatibility conditions for the finite-volume joint densities. These conditions take the form of nonlinear integral equations for boundary laws and characterize the splitting annealed Gibbs measures constructed here. For $k=1$, the translation-invariant equations reduce to positive compact integral operators. Under the stated hypotheses, they have a unique normalized positive solution. For $k=2$, an explicit empirical-loss kernel reduces the translation-invariant equations to polynomial equations. In this case, we show that hierarchical learning systems may exhibit phase-transition phenomena: for certain empirical kernels on Cayley trees order two, multiple Gibbs measures on spins emerge beyond a critical inverse temperature, corresponding to distinct equilibrium prediction regimes. For a selected annealed Gibbs measure, we distinguish its marginal on interaction parameters from its conditional Gibbs measures on spins at fixed interaction parameters. Our results provide a new perspective on energy-based learning, where data do not merely determine an optimal model through minimization but define an entire probabilistic landscape of possible inference states.
cs.LG · math.PR