Hardware influence in the stability of recurrent neural networks

It is shown that the hardware implementation is important in evaluating recurrent neural networks operating in a steady-state fashion. The stability of such networks depends on the hardware on which they are to be implemented. The specific example illustrated is the Hopfield decomposition problem. In particular, the work presented starts with an analysis of the specific circuit, shows how to derive an optimal amplifier gain, and compares it with hardware implementation results. The formalism allows the circuit gain to be set such that false minima are avoided while the circuit operates in a continuous fashion