Learning method of neural network

• US 5,293,454 A
• Filed: 09/18/1991
• Issued: 03/08/1994
• Est. Priority Date: 03/08/1989
• Status: Expired due to Fees

• Current Assignee: Sharp Kabushiki Kaisha (Hon Hai Precision Industry Co., Ltd.)
• Original Assignee: Sharp Kabushiki Kaisha (Hon Hai Precision Industry Co., Ltd.)
• Inventors: Kamiya, Shin
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/761,387
• Time in Patent Office: 902 Days
• Field of Search: 395/23
• US Class Current: 706/25
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Abstract

A learning method of a neural network, in which from a set of learning patterns belonging to one category, specific learning patterns located at a region close to learning patterns belonging to another category are selected and learning of the neural network is performed by using the specific learning patterns so as to discriminate the categories from each other.

Claims

1. A method of selecting specific learning patterns from learning patterns utilized in a neural network for speech recognition or character recognition, by classifying learning patterns into types by the K (where K is an integer>

0) nearest neighbor method, comprising the steps of;

(a) dividing the learning patterns into two categories;

(b) selecting a first target learning pattern from the learning patterns of a first category;

(c) selecting K learning patterns from the learning patterns of the first category in increasing order of distance from the first target learning pattern;

(d) calculating an average Euclidean distance between the K learning patterns and the first target learning pattern for the first category;

(e) repeating steps (c)-(d) replacing the first category with a second category;

(f) comparing a ratio of the average Euclidean distances of the K learning patterns of the first and second categories with a first predetermined threshold T₁ which is less than 1, and a second predetermined threshold, T₂, which is greater than 1;

(g) classifying the first target learning pattern as type (1) if the ratio of the average Euclidean distances of the K learning patterns of the first and second categories is between T₁ and T₂ ;

(h) selecting the first target learning pattern from the first category for utilization in the neural network for speech recognition or character recognition if it is type (1);

(i) selecting a second target learning pattern from the learning patterns of the second category;

(j) repeating steps (c)-(g) replacing the first target learning pattern with the second target learning pattern; and

(k) selecting the second target learning pattern from the second category for utilization in the neural network for speech recognition or character recognition if it is type (1).

2. The method of selecting specific learning patterns of claim 1, wherein learning patterns are classified into types using a discriminating function.

3. The method of selecting specific learning patterns of claim 1, wherein steps (d) and (f) are performed using an average city block distance calculation.

4. The method of selecting specific learning patterns of claim 1, wherein the neural network is multilayered perception type neural network.

5. The method of selecting specific learning patterns of claim 1, wherein the neural network is a Kohonen type neural network.