05276771 is referenced by 46 patents and cites 51 patents.

A data processing system and method for solving pattern classification problems and function-fitting problems includes a neural network in which N-dimensional input vectors are augmented with at least one element to form an N+j-dimensional projected input vector, whose magnitude is then preferably normalized to lie on the surface of a hypersphere. Weight vectors of at least a lowest intermediate layer of network nodes are preferably also constrained to lie on the N+j-dimensional surface.

To train the network, the system compares network output values with known goal vectors, and an error function (which depends on all weights and threshold values of the intermediate and output nodes) is then minimized. In order to decrease the network's learning time even further, the weight vectors for the intermediate nodes are initially preferably set equal to known prototypes for the various classes of input vectors. Furthermore, the invention also allows separation of the network into sub-networks, which are then trained individually and later recombined. The network is able to use both hyperspheres and hyperplanes to form decision boundaries, and, indeed, can converge to the one even if it initially assumes the other.

Title
Rapidly converging projective neural network
Application Number
7/814357
Publication Number
5276771
Application Date
December 27, 1991
Publication Date
January 4, 1994
Inventor
Gregg D Wilensky
Venice
CA, US
Narbik Manukian
Glendale
CA, US
Agent
Poms Smith Lande & Rose
Assignee
R & D Associates
CA, US
IPC
G06F 15/16
View Original Source