×

Self organizing adaptive replicate (SOAR)

  • US 5,598,510 A
  • Filed: 10/18/1993
  • Issued: 01/28/1997
  • Est. Priority Date: 10/18/1993
  • Status: Expired due to Fees
First Claim
Patent Images

1. A neural network comprising:

  • a plurality of input nodes each receiving one or more input node inputs and producing an input node output;

    a plurality of hidden nodes each receiving one or more hidden node inputs and producing a hidden node output, each of said hidden node outputs being a non-linear function of their respective hidden node inputs;

    one or more output nodes, each receiving one or more output node inputs and producing an output node output;

    first set of weighted connections coupling said input node outputs to said hidden nodes, the weights of said connections for each hidden node comprising an input weight vector;

    second set of weighted connections coupling said hidden node outputs to said output nodes, the weights of said connections for each output node comprising a hidden node weight vector, wherein substantially every input node is connected to every hidden node, and substantially every hidden node is connected to every output node;

    means for computing a set of first distance vectors that are a function of the difference between each input weight vector and an input feature vector comprising the inputs to said input nodes, said first distance vectors being fed to the hidden nodes as inputs;

    means for computing a set of second distance vectors that are a function of the difference between each hidden node weight vector, and a vector comprising the outputs of said hidden nodes;

    means for determining the smallest of the second distance vectors;

    means for generating an output from the output node associated with the smallest second distance vector, whereby for each unique input feature vector one output node generates a response and;

    means for training said network in response to a training feature vector including, means for determining if the smallest of said first distance vectors is less than a predetermined threshold and,means for adjusting the weights for the hidden node weight vector associated with said smallest first distance vector so as to make said distance smaller if the smallest of said first distance vectors is less than the predetermined threshold.

View all claims
  • 2 Assignments
Timeline View
Assignment View
    ×
    ×