On the Design of an MIMD Neural Network Processor

The design of a unique MIMD Neural Network Processor is described. The MIMD parallel processing architecture chosen facilitates the implementation of a wide variety of neural network paradigms with maximum efficiency. This efficiency is achieved by using an instruction set which is optimized for neural network processing allowing one to compute a neuron activation without arranging the weight matrix into linear arrays and/or inserting “artificial zero weighted connections”, using an MIMD parallel processing architecture to permit neurons with totally different input topologies to be updated simultaneously without loss of efficiency, and using dual neuron memories to virtually eliminate memory contention and maintain absolute memory coherence.