Finite size effects in neural networks
In this paper we give an overview of a recently developed theory [1, 2] which allows for calculating finite size corrections to the dynamical equations describing the dynamics of separable Neural Networks, away from saturation. According to this theory, finite size effects are described by a linear-noise Fokker Planck equation for the fluctuations (corresponding to an Ornstein-Uhlenbeck process), whose solution is characterized by the first two moments. The theory is applied to a particular problem in which detailed balance does not hold.
Unable to display preview. Download preview PDF.