A mesh topology for programmable neural computing

A nonfeedforward artificial neural network is simulated using concurrent processes. Reaction-diffusion neurons are used instead of Adaline neurons. To evolve the mesh architecture, a decentralized learning algorithm is used. Each neuron is individually programmed through interaction with its immediate neighbors. A `copy thy neighbor´ rule augmented with random mutations is utilized. Experiences with the algorithm indicate that such random mutations are necessary to surpass the performance of best neighbors. In addition, some of the problems being tackled by backpropagation techniques are eliminated since there are no hidden layers. The concurrent processing more closely reflects the highly parallel computational mode exhibited by living organisms. The solution of simple pattern recognition tasks with the network is described. The performance of the network compares favorably with that of conventional, sequentially simulated feedforward connectionist networks