Nonequilibrium thermodynamics of input-driven networks

TL;DR

This paper explores the nonequilibrium thermodynamics of neural networks driven by time-varying inputs, linking physical principles to neural computation and biological plausibility.

Contribution

It characterizes the thermodynamic behavior of driven neural networks, applying fluctuation theorems to understand nonequilibrium neural dynamics.

Findings

Work distributions match equilibrium free energy differences.

Fluctuation theorems apply to neural network trajectories.

Biological constraints influence nonequilibrium thermodynamics.

Abstract

Neural dynamics of energy-based models are governed by energy minimization and the patterns stored in the network are retrieved when the system reaches equilibrium. However, when the system is driven by time-varying external input, the nonequilibrium process of such physical system has not been well characterized. Here, we study attractor neural networks, specifically the Hopfield network, driven by time-varying external input and measure thermodynamic quantities along trajectories between two collective states. The overlap between distribution of the forward and reversal work along the nonequilibrium trajectories agrees with the equilibrium free energy difference between two states, following the prediction of Crooks fluctuation theorem. We study conditions with different stimulation protocol and neural network constraints. We further discuss how biologically plausible synaptic connections and information processing may play a role in this nonequilibrium framework. These results demonstrate how nonequilibrium thermodynamics can be relevant for neural computation and connect to recent systems neuroscience studies with closed-loop dynamic perturbations.