VLSI Hopfield Networks

Abstract

The Hopfield networks [1] are very popular for electronic neuro-computing due to the simplicity in the network architecture and the fast convergence property. A Hopfield network composed of one-layer neurons and fully connected feedback synapses can be used to realize associative memory, pattern classifier, and optimization circuits. The network always operates along a decreasing direction for the energy function, so that the final output represents one minimum of the energy function. Due to the complexity of energy function, there could exist several local minima. The minima in the energy function of Hopfield networks, which are encoded in the resistive network, are used for the exemplar patterns in associative memory and pattern classifier applications. However, the existence of local minima is not desirable for a great variety of other optimization applications.