A neural network to design neural networks
The design of the Hopfield associative memory is reformulated in terms of a constraint satisfaction problem, and an electronic neural net capable of solving this problem in real time is proposed.
Abstract
The design of the Hopfield associative memory is reformulated in terms of a constraint satisfaction problem. An electronic neural net capable of solving this problem in real time is proposed. Circuit solutions correspond to symmetrical zero-diagonal matrices that possess few spurious stable states. The stability of the net is proved using a suitable Lyapunov function, and simulation results are presented. The proposed network also permits design of an associative memory with a given set of state transitions, avoiding the computation of pseudo-inverses. The net exhibits several features that make it attractive for VLSI implementation. >