Link between energy and computation in a physical model of Hopfield network

Linking information processing and energy flows via thermodynamics, Landauer (1961) proposed that irreversible computational processes have an inevitable "thermodynamic cost". We explore the existence of such a link in case of a neural network model of associative memory. Our simulations with an electronic implementation of the Hopfield neural network showed that enhanced performance of the network could only be obtained by increased dissipation of energy as heat. Contrarily, efforts to minimize energy dissipation led to impaired performance.