Generalized artificial neural networks (GANN)

Based on various styles of published artificial neural networks (ANN), yet inspired by our desire to employ those aspects of neuro-computation theory which are more compatible with current VLSI implementation techniques, we describe a general model for neural networks with localized storage parameters. As a special case covered by the model introduced, we employ a quadratic relation similar to that found in practical MOS devices to implement synapses in ANN. A simulator is developed to account for the activity of synapses and to apply additional constraints for bounded weights. Application of one particular version of quadratic synapses has been examined, and positive results are shown both for the simulation of logical functions of two variables, and the detection of characters on a 3×5 retina. Also a geometrical interpretation is derived, which emphasizes that this approach has significant advantages if utilized in nonlinear pattern-recognition problems regardless of its hardware implementation convenience