VLSI implementable associative memory based on neural network architecture
Winston Ki-Cheong Mok · cIRcle (University of British Columbia) · 2010
Neural network associative memories based on Hopfield's design have two failure modes. 1) Certain memories are not recallable as they do not lie in a local minimum of the system's Liapunov (energy) function, and 2) the associative memory converges to a non-memory location due to trapping by spurious local minima in the energy function surface. The properties of the first failure mechanism in the Hopfield and five related models were investigated. A new architecture eliminating such failures is proposed. The architecture is fully digital and modular. Furthermore, it is more silicon-area efficient than any of the analyzed models. VLSI circuits built using this architecture will be able to contain a large number of neurons and several circuits may be connected together to further increase capacity.