Neural network pattern recognition learning method

• US 5,317,675 A
• Filed: 06/27/1991
• Issued: 05/31/1994
• Est. Priority Date: 06/28/1990
• Status: Expired due to Fees

• Current Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Original Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Inventors: Ikehara, Tadashi
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): DAVIS, GEORGE B
• Application Number:
  US07/722,295
• Time in Patent Office: 1,069 Days
• Field of Search: 395/23, 395/24, 395/25
• US Class Current: 706/25
• CPC Class Codes: G06F 18/24143 Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]

Abstract

A neural network includes an input layer composed of a plurality of cells receiving respective components of an input vector, an output layer composed of a plurality of cells representing attribute of the input vector, and an intermediate layer composed of a plurality of cells connected to all the cells of the input and output layers for producing a mapping to map a given input vector to its correct attribute. A learning method utilizing such neural network is carried out by image projecting the input vector into the partial dimensional space by a projection image operating means preliminarily prepared and by storing a coupling vector on the image projection space as well as the threshold and attribute vector.

Claims

1. A method of classifying a pattern in a neural network, the method comprising the steps of:

inputting an input signal into an input layer of said network, said input layer having input cells in a plurality at least as great as a number of dimensions of an input vector represented by said input signal;

transmitting said input signal to each cell in an intermediate layer, each said cell in said intermediate layer storing at least a partial dimensional space of said input vector;

activating in varying degrees those of said intermediate cells having a predetermined partial dimensional space corresponding in said varying degrees to said input vector, thereby projecting said input signal to said predetermined partial dimensional spaces and setting an activation value for each said intermediate cell; and

transmitting to each output cell in an output layer of said network each activation value weighted by a predetermined attribute vector defining coupling between each said intermediate cell and each said output cell, wherein a said input vector, I, is projected onto a said predetermined partial dimensioned space by an operator G₁, such that an image projection I'"'"'=G₁ .I and said activation value R₁ for a first intermediate cell is;

##EQU4## and wherein W₁ is a coupling vector, and ξ

_l is a threshold.

2. A method of learning patterns in a neural network, the method comprising the steps of:

routing an input vector signal to input cells in an input layer and routing said input signal to each intermediate cell in an intermediate layer;

at each said intermediate cell applying a stored image projection operator for said intermediate cell to said input signal, thereby projecting said input signal to a partial dimensional space corresponding to said intermediate cell to arrive at an image projection of said input signal on each said intermediate cell;

from each said image projection, determining an activation value for each intermediate cell for said input signal;

activating each output cell in an output layer as an average of products of said activation value and an attribute vector defining coupling between each said intermediate cell and each said output cell to arrive at an output vector signal;

comparing said output vector signal to a teacher vector signal to obtain an error signal having components; and

if a component of said error signal is not within an error tolerance, E, then

i) producing a new intermediate layer cell if no intermediate layer cell is activated by said input signal; and

ii) if a said intermediate layer cell forming said error is activated, reducing a threshold so as not to activate said intermediate layer cell forming said error with said input signal.

3. The method as recited in claim 2, wherein said new intermediate layer cell stores said input signal as said activation value, said teacher vector signal as said attribute vector, and said output vector signal as said image projection operator.