US 6,983,265 B2
Method to improve the data transfer rate between a computer and a neural network
Pascal Tannhof, Fontainebleau (France); and Ghislain Imbert de Tremiolles, Vence (France)
Assigned to International Business Machines Corporation, Armonk, N.Y. (US)
Filed on Dec. 10, 2002, as Appl. No. 10/316,250.
Claims priority of application No. 01480142 (EP), filed on Dec. 21, 2001.
Prior Publication US 2003/0135475 A1, Jul. 17, 2003
Int. Cl. G06N 3/06 (2006.01)
U.S. Cl. 706—26 5 Claims
OG exemplary drawing
 
1. A method for improving the data transfer rate of input patterns, referred to as vectors, between a host computer and an artificial neural network (ANN) implemented in hardware during a recognition phase comprising the steps of:
a) providing a set of vectors (Un, . . . ) having q components;
b) merging a set of k consecutive vectors into a single vector having p components, referred to as a base consolidated vector (U′*n), configured to globally represent said set of vectors, said base consolidated vector having only once all the components of the vectors, the first q components being formed by the components of first vector of said vector set, the (p−q) components consisting of the last component of the remaining vectors;
c) providing an artificial neural network (ANN) configured as a combination of k sub-networks operating in parallel, each neuron storing at least p components;
d) creating a set of k consolidated vectors (U′*n, . . . ,) having p components derived from said base consolidated vector, wherein each consolidated vector has q components in common with the base consolidated vector, the (p−q) remaining components being assigned a don't care condition, the first q components of the first consolidated vector having a direct correspondence with the first q components of the base consolidated vector, with the remaining components remaining as don't care, and repeating the process until the last consolidated vector which has its q last components in direct correspondence with the q last components of the base consolidated vector, the first (p−q) components having a don't care condition; and
e) presenting a new base consolidated vector to each sub-network of said ANN for recognition, wherein each of said sub-networks analyses in parallel a portion of said base consolidated vector.