Method for training an adaptive statistical classifier with improved learning of difficult samples

• US 5,805,730 A
• Filed: 08/08/1995
• Issued: 09/08/1998
• Est. Priority Date: 08/08/1995
• Status: Expired due to Term

• Current Assignee: Apple Inc.
• Original Assignee: Apple Computer Incorporated (Apple Inc.)
• Inventors: Yaeger, Larry S., Lyon, Richard F.
• Primary Examiner(s): SHALWALA, BIPIN H
• Application Number:
  US08/512,296
• Time in Patent Office: 1,127 Days
• Field of Search: 382/182, 382/185-188, 382/215, 382/220, 382/225, 382/227, 382/228-229, 382/224, 382/155-157, 382/159-161, 382/137, 364/276.6, 364/972.7, 395/21, 395/23, 395/761, 395/792
• US Class Current: 382/228
• CPC Class Codes: G06F 18/21 Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation

Abstract

A statistical classifier that can be used for pattern recognition is trained to recognize negative, or improper patterns as well as proper patterns that are positively associated with desired output classes. A set of training samples includes both the negative and positive patterns, and target output values for the negative patterns are set so that no recognized class is indicated. The negative patterns are selected for training with less frequency than the positive patterns, and their effect on training is also modified, so that training is focused more heavily on improper patterns.

Claims

1. A method for training a statistical classifier to estimate the probability that an input pattern is associated with a predetermined class, comprising the steps of:

defining a set of training patterns, each of which is labeled as belonging to a respective one of a plurality of predetermined classes;

assigning a probability of usage factor to said training patterns from said set for input to the classifier;

selecting individual training patterns;

selectively processing the selected training patterns in the classifier, or skipping the selected patterns, in accordance with said probability of usage factor which is based upon whether the samples have been properly classified previously;

detecting whether the classifier produces an output value which correctly identifies the class to which a processed pattern belongs; and

modifying the probability of usage factor for correctly identified patterns to be different from a probability of usage factor assigned to incorrectly identified patterns.

2. The method of claim 1 wherein the probability of usage factor assigned to correctly identified patterns is less than the probability of usage factor assigned to incorrectly identified patterns.

3. The method of claim 1 further including the steps of repeatedly processing training samples in the classifier over multiple successive iterations, and increasing the probability of usage factor for later iterations.

4. A method for training a statistical classifier to estimate the probability that an input pattern is associated with a predetermined class, comprising the steps of:

defining a set of training patterns, which are respectively associated with predetermined classes;

assigning a probability factor to said training patterns;

selecting individual training patterns from said set for input to the classifier;

for each selected pattern, selectively processing the selected training pattern in the classifier, in accordance with said probability factor which is based upon whether the sample have been properly classified previously;

detecting whether the classifier produces an output value which correctly identifies a class with which a processed pattern is associated; and

decreasing the probability factor for correctly identified patterns, relative to a probability factor assigned to incorrectly identified patterns.

5. The method of claim 4 further including the steps of repeatedly processing training samples in the classifier over multiple successive iterations, and increasing the probability factor for correctly identified patterns in later iterations.