Non-equilibrium phase transition in the Brownian Ising Model: field theory, renormalization group, and exact results

What the math gives to ML

The paper provides a non-equilibrium stochastic mechanism that is structurally different from iid dropout or isotropic optimizer noise: an autonomous conserved density diffuses while its Poissonian fluctuations modulate a separate order-parameter field. The transferable asset is the combination of conservation, state-dependent noise amplitude, and temporally white but spatially correlated fluctuations. A practical neural-network adaptation is to attach a positive density field to spatial tokens or feature-map locations and use its fluctuations to perturb residual activations or attention logits. The key test is whether this structured noise improves generalization or robustness over iid Gaussian noise and dropout at matched variance and compute.