On the Thermodynamics of the Hopfield Neural Network with External Field

Abstract

In the paper thermodynamic properties of an artificial neural network are analyzed in a way analogous to spin glasses theory. Synaptic connections are calculated numerically according to the Hebb rule and their distribution is obtained for different characteristics of stored patterns. The phase diagrams and magnetization are established in dependence on the temperature of the network and the external field (threshold). It was showed that changing control parameters typical of artificial neural network (i.e. number of stored patterns and pattern bias level) one obtains the results similar to the Sherrington-Kirkpatrick model of spin glass.