Abstract:The algorithms used to train neural networks, like stochastic gradient descent (SGD), have close parallels to natural processes that navigate a high-dimensional parameter space -- for example protein folding or evolution. Our study uses a Fokker-Planck approach, adapted from statistical physics, to explore these parallels in a single, unified framework. We focus in particular on the stationary state of the system in the long-time limit, which in conventional SGD is out of equilibrium, exhibiting persistent currents in the space of network parameters. As in its physical analogues, the current is associated with an entropy production rate for any given training trajectory. The stationary distribution of these rates obeys the integral and detailed fluctuation theorems -- nonequilibrium generalizations of the second law of thermodynamics. We validate these relations in two numerical examples, a nonlinear regression network and MNIST digit classification. While the fluctuation theorems are universal, there are other aspects of the stationary state that are highly sensitive to the training details. Surprisingly, the effective loss landscape and diffusion matrix that determine the shape of the stationary distribution vary depending on the simple choice of minibatching done with or without replacement. We can take advantage of this nonequilibrium sensitivity to engineer an equilibrium stationary state for a particular application: sampling from a posterior distribution of network weights in Bayesian machine learning. We propose a new variation of stochastic gradient Langevin dynamics (SGLD) that harnesses without replacement minibatching. In an example system where the posterior is exactly known, this SGWORLD algorithm outperforms SGLD, converging to the posterior orders of magnitude faster as a function of the learning rate.

Nonequilibrium fluctuations of a driven quantum heat engine via machine learning

Learning coherences from nonequilibrium fluctuations in a quantum heat engine

Leveraging Machine Learning to Gain Insights on Quantum Thermodynamic Entropy

Efficiency fluctuations of a heat engine with noise induced quantum coherences

Machine learning in and out of equilibrium

Noise induced coherent ergotropy of a quantum heat engine

Fluctuation based interpretable analysis scheme for quantum many-body snapshots

Machine learning stochastic differential equations for the evolution of order parameters of classical many-body systems in and out of equilibrium

Time-resolved Stochastic Dynamics of Quantum Thermal Machines

Geometry of learning neural quantum states

Universal fluctuations and noise learning from Hilbert-space ergodicity

Thermodynamic Geometry of Nonequilibrium Fluctuations in Cyclically Driven Transport

Optimal control of quantum thermal machines using machine learning

Pareto-optimal cycles for power, efficiency and fluctuations of quantum heat engines using reinforcement learning

Heuristic machinery for thermodynamic studies of SU(N) fermions with neural networks

Machine learning analysis of dimensional reduction conjecture for nonequilibrium Berezinskii-Kosterlitz-Thouless transition in three dimensions

Finite-Time Dynamics of an Entanglement Engine: Current, Fluctuations and Kinetic Uncertainty Relations

Machine learning one-dimensional spinless trapped fermionic systems with neural-network quantum states

Study of bounds on non-equilibrium fluctuations for asymmetrically driven quantum Otto engine

Finding the Dynamics of an Integrable Quantum Many-Body System via Machine Learning