Neural network model in pattern recognition using probabilistic contextual information

• US 4,876,731 A
• Filed: 02/19/1988
• Issued: 10/24/1989
• Est. Priority Date: 02/19/1988
• Status: Expired due to Term

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: NYNEX Corp. (Verizon Communications Inc.)
• Inventors: Euchner, James, Loris, Keith
• Primary Examiner(s): Moore, David K.
• Assistant Examiner(s): COUSO, YON JUNG
• Application Number:
  US07/158,215
• Time in Patent Office: 613 Days
• Field of Search: 382/40, 382/36, 382/15, 382/7, 382/14, 382/11
• US Class Current: 382/229
• CPC Class Codes: G06N 3/042 Knowledge-based neural networks; Logical representations of neural networks
• CPC Class Codes: G06V 30/274 Syntactic or semantic context, e.g. balancing

Abstract

A pattern recognition system for recognizing an unknown pattern comprised of symbols which are part of a pattern system which is devoid of inherent context such as numbers. Artificial contextual information based on other than symbol features and the pattern system and in the form of probability weighted expected interpretations are stored and used in the processing phase of recognition. In the system disclosed, the system comprises a neural network whose forward and feedback paths are controlled by the output cells of the network based, in part, on the contextual information.

Claims

1. A system for recognizing an unknown pattern, said unknown pattern including one or more symbols which are part of a pattern system which is devoid of inherent context, comprising:

storing means for storing artificial contextual information associated with said unknown pattern and based upon other than information as to symbol features and the pattern system; and

recognizer means responsive to said known pattern and to said artificial contextual information, said recognizer means having a processing phase wherein said recognizer means processes at least one symbol of said unknown pattern and outputs as a recognition decision that said one symbol is one of a plurality of category names and wherein in reaching said recognition decision during said processing phase utilizes said artificial contextual information.

2. A system in accordance with claim 1 wherein:

said unknown pattern includes a plurality of symbols;

and said recognizer means during said processing phase processes each of said symbols of said unknown pattern and outputs as a recognition decision for each of said symbols that the symbol is one of a plurality of category names and in reaching said recognition decision during said processing phase utilizes said derived contextual information.

3. A system in accordance with claim 2 wherein:

said derived contextual information comprises expected interpretations of said unknown pattern and probability values associated with such expected interpretations.

4. A system in accordance with claim 3 wherein:

said unknown pattern is a sequence of individual numbers.

5. A system in accordance with claim 4 wherein:

said numbers correspond to the dollar amount of a check.

6. A system in accordance with claim 5 wherein:

said check is utilized in payment of a bill.

7. A system in accordance with claim 6 wherein:

one of said expected interpretations of said unknown pattern corresponds to the dollar amount of said bill.

8. A system in accordance with claim 7 wherein:

others of said expected interpretations of said unknown pattern correspond to the dollar amount of a bill previous to said bill and the sum of the dollar amounts of said bill and previous bill.

9. A system in accordance with claim 8 wherein:

said bills are telephone bills.

10. A system in accordance with claim 2 wherein:

said recognizer means comprises;

neural network means responsive to said unknown pattern for carrying out said processing phase; and

selector means responsive to said derived contextual information for controlling said neural network.

11. A system in accordance with claim 10 wherein:

said neural network means comprises;

a plurality of stages of cells including an input stage for receiving said unknown pattern and an output stage having output cells each corresponding to a different one of said category names and to a different set of essential symbol features;

an afferent path connecting said stages; and

an efferent path connecting said stages for facilitating and inhibiting said afferent path; and

said selector means is responsive to said output cells and based thereon and on said derived contextual information controls said efferent path.

12. A system in accordance with claim 11 wherein:

said selector means selects one of said output cells to condition said efferent path so said efferent path facilitates transmission in said afferent path associated with the set of essential symbol features corresponding to said selected output cell.

13. A system in accordance with claim 12 wherein:

said selector means comprises a rule based selector operating on the basis of certainty factor calculus.

14. A system in accordance with claim 10 wherein:

said recognizer means comprises means for automatically changing the attention of said neural network when said recognizer means outputs a recognition decision with respect to a symbol in said pattern.

15. A system in accordance with claim 2 wherein;

said recognizer means comprises means for deleting from said unknown pattern the symbols of said unknown pattern for which said recognizer means has already made a recognition decision.

16. A system in accordance with claim 2 wherein:

said recognizer means includes means for adjusting said derived contextual information in dependence on said recognizer means reaching a recognition decision with respect to the individual symbols in said pattern.

17. A method of recognizing an unknown pattern, said unknown pattern including one or more symbols which are part of a pattern system which is devoid of inherent context, comprising:

storing artificial contextual information associated with said unknown pattern and based upon other than information as to symbol features and the pattern system; and

recognizing said unknown pattern by processing at least one symbol of said unknown pattern and outputting as a recognition decision that said one symbol is one of a plurality of category names utilizing said artificial contextual information.

18. A method in accordance with claim 17 wherein:

said unknown pattern includes a plurality of symbols;

and said recognizing includes processing each of said symbols of said unknown pattern and outputting a recognition decision for each of said symbols that the symbol is one of a plurality of category names utilizing said derived contextual information.

19. A method in accordance with claim 18 wherein:

said derived contextual information comprises expected interpretations of said unknown pattern and probability values associated with such expected interpretations.

20. A method in accordance with claim 19 wherein:

said unknown pattern is a sequence of individual numbers.

21. A method in accordance with claim 20 wherein:

said numbers correspond to the dollar amount of a check.

22. A method in accordance with claim 21 wherein:

said check is utilized in payment of a bill.

23. A method in accordance with claim 22 wherein:

one of said expected interpretations of said unknown pattern corresponds to the dollar amount of said bill.

24. A method in accordance with claim 23 wherein:

others of said expected interpretations of said unknown pattern correspond to the dollar amount of a bill previous to said bill and the sum of the dollar amounts of said bill and previous bill.

25. A method in accordance with claim 24 wherein:

said bills are telephone bills.

26. A method in accordance with claim 18 wherein:

said processing step is carried out by neural network means responsive to said unknown pattern and by selector means responsive to said derived contextual information for controlling said neural network.

27. A method in accordance with claim 26 wherein:

said neural network means comprises;

a plurality of stages of cells including an input stage for receiving said unknown pattern and an output stage having output cells each corresponding to a different one of said category names and to a different set of essential symbol features;

an afferent path connecting said stages; and

an efferent path connecting said stages for facilitating and inhibiting said afferent path; and

said selector means is responsive to said output cells and based thereon and on said derived contextual information and controls said efferent path.

28. A method in accordance with claim 27 wherein:

said selector means selects one of said output cells to condition said efferent path so said efferent path facilitates transmission in said afferent path associated with the set of essential symbol features corresponding to said selected output cell.

29. A method in accordance with claim 28 wherein:

said selector means comprises a rule based selector operating on the basis of certainty factor calculus.

30. A method in accordance with claim 26 wherein:

said recognizing comprises automatically changing the attention of said neural network when a recognition decision with respect to a symbol in said pattern is reached.

31. A method in accordance with claim 18 wherein:

said recognizing step includes deleting from said unknown pattern the symbols of said unknown pattern for which a recognition decision has already been reached.

32. A method in accordance with claim 18 wherein:

said recognizing step includes adjusting said derived contextual information in dependence on reaching a recognition decision with respect to the individual symbols in said pattern.