Classification method implemented in a layered neural network for multiclass classification and layered neural network

• US 5,220,618 A
• Filed: 02/08/1991
• Issued: 06/15/1993
• Est. Priority Date: 02/09/1990
• Status: Expired due to Fees

• Current Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Original Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Inventors: Nadal, Jean-Pierre, Sirat, Jacques-Ariel
• Primary Examiner(s): Boudreau, Leo H.
• Application Number:
  US07/653,592
• Time in Patent Office: 858 Days
• Field of Search: 382/14, 382/15, 382/36, 395/21, 395/23, 395/24
• US Class Current: 382/157
• CPC Class Codes: G06F 18/24137 Distances to cluster centroïds
• CPC Class Codes: G06F 18/24323 Tree-organised classifiers
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology

Abstract

Classification method implemented in a layered neural network, comprising learning steps during which at least one layer is constructed by the addition of the successive neurons necessary for operating, by successive dichotomies, a classification of examples distributed over classes. In order to create at least one layer starting with a group of examples distributed over more than two classes, each successive neuron tends to distinguish its input data according to two predetermined sub-groups of classes peculiar to the said neuron according to a principal components analysis of the distribution of the said input data subjected to the learning of the neuron of the layer in question.

Claims

1. A method of classifying a group of non-homogenous data into K homogenous classes by training a neural network device, said device including at least one processing element, having a classification coefficient, comprising the steps of:

a) dividing said group of non-homogenous data into two subgroups of data;

b) applying one of said two subgroups of data to said processing element to generate a classification coefficient for said processing element;

c) applying said group of non-homogeneous data to said processing element;

d) processing said group of non-homogeneous data into two further groups of data in accordance with the division of said two subgroups of data in step a);

e) testing the homogeneity of each of said two further groups to determine if each of said further subgroups contains only a single class;

f) storing a location of said processing element and recording said homogeneous class of one of said further groups of data when said processing element distinguishes said single class;

g) if one of said further groups is non-homogeneous, generating a second processing element; and

h) repeating steps a)-d) until said further groups of data are classified by said successive processing element so that each further group corresponds to a homogeneous class.

3. The method of claim 1, wherein the step of dividing is performed in accordance with a principal components analysis.

5. The method of claim 1, wherein said network parameters are calculated using one of a Pocket algorithm, a Perceptron learning algorithm and a Hebb algorithm.

6. The method of claim 1, wherein K is greater than 2.

2. A method of classifying non-homogeneous data into more than two classes of homogeneous data by training a neural network, said network including a plurality of processing elements, said processing elements having at least one network parameter stored in an interconnection between the processing elements that specify weights of the interconnections, said method comprising the steps of:

a.) dividing the non-homogeneous data into a first subgroup of data and a second subgroup of data in accordance with a principal components analysis;

b.) applying the non-homogeneous data to a first processing element and propagating in a forward propagating step until generation of a first group of result data and second group of result data;

b.) adjusting the network parameters in accordance with the principal components analysis of said first subgroup of data and said second subgroup of data;

c.) determining whether the first group of result data and the second group of result data are non-homogeneous and wherein said at least one of said first group of result data and said second group of result data is non-homogeneous, generating a further processing element coupled to said processing element; and

d.) repeating steps a)-c) with respect to the first and second groups of result data and said further processing element until each of said subgroups is homogeneous.

4. The method of claim 2, wherein said network parameters are calculated using one of a Pocket algorithm, a Perceptron learning algorithm and a Hebb algorithm.

7. A neural network device for classifying a group of non-homogeneous data into four homogeneous classes, comprising:

an input for receiving said non-homogeneous data;

a first processing element connected to said input, said processing element using a synaptic coefficient for separating said group of non-homogeneous data into two further groups of non-homogeneous data;

a second processing element coupled to said first processing element, said second processing element for receiving said two further groups of non-homogeneous data and separating said two further groups of non-homogeneous data into four groups of homogeneous data;

means for adjusting the network parameters of each of said processing elements; and

means for validating whether an output of first and second processing elements is homogeneous data.