Data learning system for identifying, learning apparatus, identifying apparatus and learning method

• US 7,783,581 B2
• Filed: 12/26/2005
• Issued: 08/24/2010
• Est. Priority Date: 01/05/2005
• Status: Active Grant

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Satou, Atsushi, Hosoi, Toshinori
• Primary Examiner(s): Holmes; Michael
• Application Number:
  US11/794,730
• Publication Number:
  US 20080091627A1
• Time in Patent Office: 1,702 Days
• Field of Search: 706/12
• US Class Current: 706/12
• CPC Class Codes: G06F 18/28 Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06V 10/772 Determining representative reference patterns, e.g. averaging or distorting patterns; Generating dictionaries

Abstract

A discriminative data learning system provided comprises a learning device for updating dictionary data having a template to be used for discriminating the category of a hierarchical structure, and a discriminating device for classifying a pattern into a category with reference to the dictionary data. The learning device includes a learning data storage unit for storing the pattern and the category number of the category, to which the pattern belongs, a dictionary data storage unit for holding the template and the category number, to which the template belongs, a similarity calculating unit for calculating the similarity between the pattern and the template, an update calculating unit for calculating a first update for updating the template, on the basis of the similarity, a weighing unit for multiplying the first update by a weighing constant each determined separately for the hierarchy of the category, to determine a second update, and a dictionary updating unit for updating the template on the basis of the second update. The discriminating device includes a category determining unit for classifying the input pattern having the calculated similarity, into the same category as that of the template having a high similarity.

Claims

1. A system comprising:

a data storage section configured to hold a first pattern and templates each of which belongs to members of a set of categories, wherein each of said categories corresponds to one of hierarchies of a hierarchical structure;

a similarity calculating section configured to calculate first similarities between said first pattern and said templates;

an update calculating section configured to calculate a first update based on said first similarities;

a weighting section configured to multiply said first update by one of weight constants to provide a second update, wherein said weight constants correspond to said hierarchies, respectively;

an updating section configured to update said templates based on said second update; and

a category determining section,

wherein said similarity calculating section is further configured to calculate second similarities between a second pattern and said updated templates, and

said category determining section is configured to classify said second pattern into one of said categories based on said second similarities.

2. The system according to claim 1, further comprising:

a feature extracting section configured to convert said first pattern and said second pattern into feature patterns and to send said feature patterns to said similarity calculating section.

3. A learning apparatus, comprising;

a data storage section configured to hold a pattern and templates each of which belongs to members of a set of categories, wherein each of said categories corresponds to one of hierarchies of a hierarchical structure;

a similarity calculating section configured to calculate similarities between said pattern and said templates;

an update calculating section configured to calculate a first update based on said similarities;

a weighting section configured to multiply said first update by one of weight constants to provide a second update, wherein said weight constants correspond to said hierarchies, respectively; and

an updating section configured to update said templates based on said second update.

4. The learning apparatus according to claim 3, wherein said data storage section is further configured to hold a main data in which each of said templates is correlated to one of said members, and

said one of said members corresponds to a first hierarchy of said hierarchies.

5. The learning apparatus according to claim 4, wherein said data storage section is further configured to hold a correspondence table in which said one of said members is correlated to another of said members,

said another of said members corresponds to a second hierarchy of said hierarchies, and

said second hierarchy is other than said first hierarchy.

6. The learning apparatus according to claim 3, wherein said data storage section is further configured to hold a main data in which each of said templates is correlated to each of said members.

7. The learning apparatus according to claim 3, further comprising:

a feature extracting section configured to convert said pattern into feature pattern and to send said feature pattern to said similarity calculating section.

8. A learning apparatus comprising:

a similarity calculating section configured to calculate similarities between a pattern and templates each of which belongs to members of a set of categories, wherein each of said categories corresponds to one of hierarchies of a hierarchical structure;

an update calculating section configured to calculate a first update based on said similarities;

a weighting section configured to multiply said first update by one of weight constants to provide a second update, wherein said weight constants correspond to said hierarchies, respectively; and

an updating section configured to update said templates based on said second update.

9. The learning apparatus according to claim 8, further comprising:

a feature extracting section configured to convert said pattern into feature pattern and to send said feature pattern to said similarity calculating section.

10. An identifying apparatus comprising:

a dictionary data storage section configured to hold updated templates prepared by a learning apparatus;

an identifying apparatus similarity calculating section; and

a category determining section,

wherein said learning apparatus comprises;

a learning data storage section configured to hold a learning pattern and templates each of which belongs to members of a set of categories, wherein each of said categories corresponds to one of hierarchies of a hierarchical structure;

a learning apparatus similarity calculating section configured to calculate first similarities between said learning pattern and said templates;

an update calculating section configured to calculate a first update based on said first similarities;

a weighting section configured to multiply said first update by one of weight constants to provide a second update; and

an updating section configured to update said templates based on said second update to provide said updated template,

said weight constants correspond to said hierarchies, respectively,

said identifying apparatus similarity calculating section is configured to calculate second similarities between an input pattern and said updated templates, and

said category determining section is configured to classify said input pattern into one of said categories based on said second similarities.

11. The identifying apparatus according to claim 10, wherein a similarity between one of said updated templates and said input pattern is highest among said second similarities,

said one of said updated templates belongs to said one of said categories.

12. An identifying apparatus comprising:

a dictionary data storage section configured to hold updated templates prepared by a learning apparatus;

an identifying apparatus similarity calculating section; and

a category determining section,

wherein said learning apparatus comprises;

a learning apparatus similarity calculating section configured to calculate first similarities between a learning pattern and templates each of which belongs to members of a set of categories, wherein each of said categories corresponds to one of hierarchies of a hierarchical structure;

an update calculating section configured to calculate a first update based on said first similarities;

a weighting section configured to multiply said first update by one of weight constants to provide a second update; and

an updating section configured to update said templates based on said second update to provide said updated template,

said weight constants correspond to said hierarchies, respectively,

said identifying apparatus similarity calculating section is configured to calculate second similarities between an input pattern and said updated templates, and

said category determining section is configured to classify said input pattern into one of said categories based on said second similarities.

13. The identifying apparatus according to claim 12, wherein a similarity between one of said updated templates and said input pattern is highest among said second similarities,

said one of said updated templates belongs to said one of said categories.

14. A learning method comprising:

calculating a similarity between one of patterns and each of templates, wherein each of said patterns and said templates belongs to members of a set of categories, and said categories correspond to hierarchies of a hierarchical structure;

calculating an update based on said similarity by considering one of said hierarchies;

weighting said update by using a weight constant corresponding to said one of said hierarchies; and

updating said templates based on said weighted update.

15. The learning method according to claim 14, wherein said calculating said similarity, said calculating said update, said weighting and said updating are executed for each of said hierarchies.

16. The learning method according to claim 15, wherein said execution of said calculating said similarity, said calculating said update, said weighting and said updating is executed for each of said patterns.

17. The learning method according to claim 16, wherein said execution of said execution of said calculating said similarity, said calculating said update, said weighting and said updating is repeated for predetermined number.

18. A learning method comprising:

calculating similarities between a first pattern and templates each of which belongs to members of a set of categories, wherein each of said categories corresponds to one of hierarchies of a hierarchical structure;

calculating an update based on said similarities;

weighting said update by using one of weight constants, wherein said weight constants correspond to said hierarchies, respectively; and

updating said templates based on said weighted update.

19. The learning method according to claim 18, further comprising:

converting a second pattern into said first pattern by feature extraction.

20. The learning method according to claim 18, wherein in said calculating said similarities, a main data is referred in which each of said templates is correlated to a one of said members, and

said one of said members corresponds to a first hierarchy of said hierarchies.

21. The learning method according to claim 20, wherein in said calculating said similarities, a correspondence table is referred in which said one of said members is correlated to another of said members,

said another of said members corresponds to a second hierarchy of said hierarchies, and

said second hierarchy is other than said first hierarchy.

22. The learning method according to claim 18, wherein in said calculating said similarities, a main data is referred in which each of said templates is correlated to each of said members.

23. An identifying method comprising:

calculating a first similarity between one of patterns and each of templates, wherein each of said patterns and said templates belongs to members of a set of categories, and said categories correspond to hierarchies of a hierarchical structure;

calculating an update based on said first similarity by considering one of said hierarchies;

weighting said update by using a weight constant corresponding to said one of said hierarchies;

updating said templates based on said weighted update;

calculating similarities between an input pattern and said updated templates; and

classifying said input pattern into one of said categories based on said similarities.

24. The identifying method according to claim 23, wherein a second similarity between one of said updated templates and said input pattern is highest among said similarities,

said one of said updated templates belongs to said one of said categories.

25. An identifying method comprising:

calculating first similarities between a pattern and templates each of which belongs to members of a set of categories, wherein each of said categories corresponds to one of hierarchies of a hierarchical structure;

calculating an update based on said first similarities;

weighting said update by using one of weight constants, wherein said weight constants correspond to said hierarchies, respectively;

updating said templates based on said weighted update;

calculating second similarities between an input pattern and said updated templates; and

classifying said input pattern into one of said categories based on said second similarities.

26. The identifying method according to claim 25, wherein a similarity between one of said updated templates and said input pattern is highest among said a second similarities,

said one of said updated templates belongs to said one of said categories.

27. A computer program product for a method which comprises:

calculating a similarity between one of patterns and each of templates, wherein each of said patterns and said templates belongs to members of a set of categories, and said categories correspond to hierarchies of a hierarchical structure;

calculating an update based on said similarity by considering one of said hierarchies;

weighting said update by using a weight constant corresponding to said one of said hierarchies; and

updating said templates based on said weighted update.

28. A computer program product for a method which comprises:

calculating similarities between a pattern and templates each of which belongs to members of a set of categories, wherein each of said categories corresponds to one of hierarchies of a hierarchical structure;

calculating an update based on said similarities;

weighting said update by using one of weight constants, wherein said weight constants correspond to said hierarchies, respectively; and

updating said templates based on said weighted update.

29. A computer program product for a method which comprises:

calculating a similarity between one of patterns and each of templates, wherein each of said patterns and said templates belongs to members of a set of categories, and said categories correspond to hierarchies of a hierarchical structure;

calculating an update based on said similarity by considering one of said hierarchies;

weighting said update by using a weight constant corresponding to said one of said hierarchies;

updating said templates based on said weighted update;

calculating similarities between an input pattern and said updated templates; and

classifying said input pattern into one of said categories based on said similarities.

30. A computer program product for a method which comprises:

calculating first similarities between a pattern and templates each of which belongs to members of a set of categories, wherein each of said categories corresponds to one of hierarchies of a hierarchical structure;

calculating an update based on said first similarities;

weighting said update by using one of weight constants, wherein said weight constants correspond to said hierarchies, respectively;

updating said templates based on said weighted update;

calculating second similarities between an input pattern and said updated templates; and

classifying said input pattern into one of said categories based on said second similarities.

31. A computer readable recording media which records a dictionary data including:

a main data in which each of templates is correlated to a first category corresponding to a first hierarchy of a hierarchical structure; and

a correspondence table in which said first category is correlated to a second category corresponding to a second hierarchy of said hierarchical structure,

wherein said second hierarchy is other than said first hierarchy.

32. A computer readable recording media which records a dictionary data including:

a main data in which each of templates is correlated to a category corresponding to each hierarchy of a hierarchical structure.