Moment-Resolved Readout and Reservoir Diversity in Nonequilibrium Langevin Computing

What the math gives to ML

The paper offers a transferable readout mechanism for stochastic or recurrent neural systems: preserve multiple nonequilibrium response moments instead of compressing each state distribution to its mean. For quartic Langevin units, second- and fourth-order moments capture input-dependent width and tail changes that are invisible to mean-only readout. A second mechanism is heterogeneous reservoir fusion, where independently initialized or trained reservoirs expose complementary features before class-logit compression. The most useful neural-network test is a moment-augmented stochastic recurrent model with controlled noise and forcing, together with ablations that determine whether non-Gaussian moment information predicts gains rather than merely increasing feature count.