Extended - input neural network applications, implementation and learning

  • Authors:

  • Wafik A. Wassef

  • Affiliations:

  • Department of Computer Engineering, Saskatchewan Institute of Applied Science and Technology, Palliser Campus, Moose Jaw, Saskatchewan, P.O. Box 1420, S6H 5R4 Canada

  • Venue:
  • Neural, Parallel & Scientific Computations
  • Year:
  • 2002

Quantified Score

Hi-index 0.00

Visualization

Abstract

By adding extra input nodes and connecting them through nonlinear elements to the original inputs, it became possible to realize a neural network for any desired output. All digital logic functions are produced by this simple circuit with minimum use of nonlinear circuits. The extension of this network to any number of input nodes is given as well. A direct method for writing down the exact inverse input matrix for this network, without any calculations is described. The implementation of the linear weights and their polarity is demonstrated. Some of the applications given are: adding and multiplying two 2-bit binary numbers, the square function and an encoder that assigns a single output to each of the input states. A learning technique for producing the weights between input and output of this encoder network and its analogy to biological chemical synapses between neurons in the hippocampus is discussed.