Learning and recognition machine

• US 5,479,570 A
• Filed: 10/04/1993
• Issued: 12/26/1995
• Est. Priority Date: 10/06/1992
• Status: Expired due to Fees

• Current Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Imagawa, Taro, Kojima, Yoshihiro, Maruno, Susumu, Kouda, Toshiyuki, Shimeki, Yasuharu
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Downs, Robert W.
• Application Number:
  US08/130,851
• Time in Patent Office: 813 Days
• Field of Search: 395/11, 395/21, 395/22, 395/23, 395/24, 395/27, 395/20, 382/39, 382/14
• US Class Current: 706/20
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06V 30/242 Division of the character sequences into groups prior to recognition; Selection of dictionaries

Abstract

A learning and recognition machine has a major classification section for classifying an input pattern into a plurality of category groups and calculating a group belongingness to each category group and a plurality of fine classification sections for calculating a similarity within each category group. The machine causes the plurality of fine classification sections to learn in a coordinated manner by reflecting the group belongingness of the input pattern to update of a weight factor implemented in the fine classification sections, thereby allowing a recognition rate of a pattern situated at the boundary of the category groups to be improved and an effective supplemental learning to be performed.

Claims

1. A learning and recognition machine, comprising:

a major classification section for calculating a group belongingness which indicates a degree of how close an input pattern signal belongs to a category group composed of a set of similar categories,

a plurality of fine classification sections each having an in-group similarity calculating section for calculating an in-group similarity which indicates a degree of how close said input pattern signal belongs to a category in each category group and a weight factor updating section for updating weight factors of said in-group similarity calculating section;

a discrimination signal loading section for weighting each said in-group similarity by the group belongingness;

a category discriminating section for comparing the weighted in-group similarities output from said discrimination signal loading section;

a teacher signal generating section for generating a teacher signal output necessary for learning;

a learning control section for outputting a learning control signal for controlling weights of each said in-group similarity calculating section based on the teacher signal output from said teacher signal generating section and based on the output of said discrimination signal loading section or of said category discriminating section; and

a learning control signal loading section for weighting the learning control signal output from said learning control section by the group belongingness output from said major classification section; and

each said weight factor updating section for updating the weight factor of each of said in-group similarity calculating sections being based on the outputs of said learning control signal loading section and of each said in-group similarity calculating section.

2. The learning and recognition machine according to claim 1, wherein

said learning control section comprises:

a comparator for comparing the teacher signal output from said teacher signal generating section and a discrimination result of said category discriminating section and

a weight modification control section for determining a learning control signal in correspondence to the comparison result of said comparator.

3. The learning and recognition machine according to claim 1, wherein

said learning control section comprises:

an error calculating section for calculating a difference of the teacher signal output from said teacher signal generating section and the output of said discrimination signal loading section ,and

a weight modification control section for determining a learning control signal in correspondence to the output of said error calculating section.

4. The learning and recognition machine according to claim 3, further comprising:

a category memory section for storing types of categories belonging to each category group;

a category adding section for adding a means for calculating a similarity to a new category to said in-group similarity calculating section in each of said fine classification sections based on the memory contents of said category based on the memory section and the teacher signal; and

an error dependent category adding control section for switching per category group whether said category adding section should add a category or not based on the error signal output by said error calculating section.

5. The learning and recognition machine according to claim 4, wherein

said error dependent category adding control section comprises:

a memory section for storing a criterion error value,

a comparator for comparing the output of said error calculating section with the error stored in said memory section, and;

a category adding directing section for directing said category adding section to add a category when the error exceeds the criterion error value.

6. The learning and recognition machine according to claims 1 further comprising:

a group selecting section for selecting a plurality of category groups based on the group belongingness output by the major classification section and

an input signal selecting section for selecting a fine classification section to which the input pattern signal is input based on the selected result of said group selecting section.

7. The learning and recognition machine according to claim 6, further comprising:

a category memory section for storing types of categories belonging to each category group,

a limited group selecting section for selecting a plurality of category groups based on memory contents of said category memory section and the group belongingness output by said major classification section and

a group selecting section switching section for switching said group selecting section and said limited group selecting section.

8. The learning and recognition machine according to claims 7, further comprising:

a group belongingness loading section for weighting the group belongingness based on the memory contents of said category memory section.

9. The learning and recognition machine according to claim 1, further comprising:

a category memory section for storing types of categories belonging to each category group;

a category adding section for adding a means for calculating a similarity to a new category to said in-group similarity calculating section in each of said fine classification sections based on the memory contents of said category memory section and the teacher signal.

10. The learning and recognition machine according to claim 9, further comprising:

a belongingness dependent category adding control section for switching per category group whether said category adding section should add category or not based on the group belongingness output by said major classification section.

11. The learning and recognition machine according to claim 10, wherein

said belongingness dependent category adding control section comprises:

a maximum value detector for finding a maximum value of the group belongingness,

a memory section for storing a criterion group belongingness,

a comparator for comparing the maximum value of the group belongingness detected by said maximum value detector with the criterion group belongingness stored in said memory section, and

a category adding directing section for directing said category adding section to add a category when the maximum value of the group belongingness exceeds the criterion group belongingness.

12. The learning and recognition machine according to claim 1, wherein

said major classification section comprises:

a group dictionary for providing a plurality of group reference pattern signals which represent category groups composed of a set of similar patterns, and

a group reference pattern similarity calculating section for calculating group belongingness using said group reference pattern signals.

13. The learning and recognition machine according to claims 12, wherein

calculation constants used in calculating the group belongingness of said major classification section can be changed.

14. The learning and recognition machine according to claim 13, further comprising:

a learning monitor section for measuring a number of input times of patterns to be learned ,and

a constant control section; and

said constant control section changing the calculation constants used in calculating the group belongingness of said major classification section based on the output of said learning monitor section.

15. The learning and recognition machine according to claim 13, further comprising:

a learning monitor section for calculating output variation of said error calculating section, and

a constant control section; and

said constant control section changing the calculation constants of said major classification section based on the output of said learning monitor section.

16. The learning and recognition machine according to claim 12, wherein

said major classification section comprises:

a plurality of distance calculating sections for calculating distances between the input pattern signal and all group reference pattern signals stored in said group dictionary,

a plurality of dividers for calculating an inverse number of an output of said distance calculating section, an adder for adding each of the outputs of said dividers,

a plurality of multipliers for multiplying the output of said adder with the outputs of said distance calculating section, and

a plurality of dividers for calculating an inverse number of the output of said multipliers.

17. The learning and recognition machine according to claim 16, wherein

said distance calculating section of said major classification section comprises:

a signal intensifier comprising a plurality of subtracters for calculating differences of a plurality of signals inputted, an adder and a plurality of involuters, and

said belongingness control section comprises a constant control section for changing values of powers of involution of said involuters in said signal intensifier.

18. The learning and recognition machine according to claims 1, wherein

said fine classification section comprises a plurality of multiinput-output signal processing sections arranged and connected in a layer structure so that signals are propagated from a lower layer to an upper layer and

said multiinput-output signal processing section comprising

a weight factor memory section for holding a plurality of weight factors,

an input section for inputting a plurality of input signals,

multiplying means for weighting the input signals from said input sections by the weight factors stored in said weight factor memory section,

adding means for adding the plurality of input signals weighted by said multiplying means, and

a threshold value processing section for limiting the output of said adding means into a certain range.

19. The learning and recognition machine according to claims 1, wherein

said fine classification section comprises:

a category dictionary in which a plurality of category reference pattern signals that indicate a central value of each category of the input pattern signal are stored,

a plurality of distance calculating sections for calculating distances between the input pattern signal and all the category reference pattern signals stored in said category dictionary,

a plurality of dividers for calculating the inverse number of the output of said distance calculating section,

an adder for adding each of the outputs of said dividers,

a plurality of multipliers for multiplying the output of said adder with the outputs of said distance calculating section and a plurality of dividers for calculating the inverse number of the output of said multipliers.

20. The learning and recognition machine according to claim 1, wherein said fine classification section is comprised in a manner that a learning and recognition unit is arranged at an upper layer of a plurality of branch structures which are constructed by combining route selecting units in a layer constitution,

each said route selecting unit comprising;

a signal input section,

a quantizer for quantizing the signal from the signal input section,

a route input section having single or plural route input terminals,

a route output section having single or plural route output terminals,

a route loading section for selecting a route in accordance to a result of the quantizing of said quantizer,

an internal state memory section for storing internal states of the route selecting unit,

a quantization control section for changing a number and a range of the quantizing of the quantizer,

a structure memorizing section for memorizing a structure of the route selecting unit, and

said learning and recognition unit comprising;

a teacher signal input section,

a route input section,

a loading section for weighting input values of the route input section,

an adder for adding outputs from the loading section, and

a learning machine for determining a load value of said loading section.

21. The learning and recognition machine according to claim 20, further comprising:

a quantization criterion memory section for storing a comparison reference to contents of the internal state memory section, and

a quantization criterion control section for changing contents of the quantization criterion memory section with respect to every fine classification section.

22. The learning and recognition machine according to claim 21, wherein

the quantization criterion control section comprises:

an error monitor section for monitoring output variation of the error calculating section, and

a quantization constant control section for changing contents stored in the quantization criterion memory section based on a result by the error monitor section.

23. The learning and recognition machine according to claim 21, wherein

the quantization criterion control section comprises:

a learning times monitor section for monitoring a number of times of inputs of pattern signal to be learned, and

a quantization constant control section for changing contents stored in the quantization criterion memory section based on a result by the learning times monitor section.

24. The learning and recognition machine according to claim 21, wherein

the quantization criterion control section comprises:

a category adding monitor section for monitoring a category adding operation of a category adding section, and

a quantization constant control section for changing contents stored in the quantization criterion memory section based on a result by the category adding monitor section.