System and method of pattern recognition employing a multiprocessing pipelined apparatus with private pattern memory

• US 5,459,798 A
• Filed: 05/12/1993
• Issued: 10/17/1995
• Est. Priority Date: 03/19/1993
• Status: Expired due to Term

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Bailey, Delbert D., Dulong, Carole
• Primary Examiner(s): Boudreau, Leo H.
• Assistant Examiner(s): Johns, Andrew W.
• Application Number:
  US08/060,579
• Time in Patent Office: 888 Days
• Field of Search: 382/13, 382/30, 382/33, 382/34, 382/41, 382/49, 364/229.2, 364/231.8, 364/240.1, 364/240.2, 364/242.3, 364/243, 364/243.43, 364/243.44, 364/243.45, 364/926.8, 364/948.34, 364/964 , 364/964.21, 364/964.26
• US Class Current: 382/218
• CPC Class Codes: F02B 2075/027 four
• CPC Class Codes: G06F 18/295 Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/285 Memory allocation or algorithm optimisation to reduce hardware requirements
• CPC Class Codes: G10L 15/34 Adaptation of a single recogniser for parallel processing, e.g. by use of multiple processors or cloud computing

Abstract

A computer implemented apparatus and method of pattern recognition utilizing a pattern recognition engine coupled with a general purpose computer system. The present invention system provides increased accuracy and performance in handwriting and voice recognition systems and may interface with general purpose computer systems. A pattern recognition engine is provided within the present invention that contains five pipelines which operate in parallel and are specially optimized for Dynamic Time Warping and Hidden Markov Models procedures for pattern recognition, especially handwriting recognition. These pipelines comprise two arithmetic pipelines, one control pipeline and two pointer pipelines. Further, a private memory is associated with each pattern recognition engine for library storage of reference or prototype patterns. Recognition procedures are partitioned across a CPU and the pattern recognition engine. Use of a private memory allows quick access of the library patterns without impeding the performance of programs operating on the main CPU or the host bus. Communication between the CPU and the pattern recognition engine is accomplished over the host bus.

Claims

1. An apparatus for performing pattern recognition procedures regarding an unknown pattern, said apparatus for interfacing to a computer system having a computer system bus, said apparatus comprising:

programmable multiprocessing means for executing pattern recognition instructions, each pattern recognition instruction comprising arithmetic, pointer and control operations which are executed in parallel, wherein said programmable multiprocessing means further comprises;

(i) two arithmetic pipelines for executing two arithmetic instructions in parallel, a first arithmetic pipeline optimized for performing distance computations between segments of said unknown pattern and segments of reference patterns and a second arithmetic pipeline optimized for performing best path computations in response to previously computed distance computations;

(ii) two pointer pipelines for executing two pointer instructions in parallel, a first pointer pipeline for supplying pointer information for said first arithmetic pipeline and a second pointer pipeline for supplying pointer information for said second arithmetic pipeline; and

(iii) a control pipeline for providing instruction flow control for said pattern recognition procedures;

internal memory means for providing storage for temporary results of said pattern recognition instructions and for providing temporary storage for said unknown pattern and selected reference patterns; and

external memory means for providing storage for a library of said reference patterns, said external memory means communicatively coupled to said internal memory means via a private and dedicated communication bus separate from said computer system bus.

2. An apparatus for performing pattern recognition procedures regarding an unknown pattern as described in claim 1 wherein said internal memory means comprises:

first data storage means coupled to said first arithmetic pipeline and coupled to said first pointer pipeline for containing said unknown pattern and for containing at least one reference pattern of said external memory means;

second data storage means coupled to said second arithmetic pipeline and coupled to said second pointer pipeline for containing distance and best path results generated from said first and said second arithmetic pipelines; and

program storage means for containing instructions executed by said programmable multiprocessing means.

3. An apparatus for performing pattern recognition procedures regarding an unknown pattern as described in claim 1 wherein said programmable multiprocessing means comprises means for interfacing with and communicating with said computer system.

4. An apparatus for performing pattern recognition procedures regarding an unknown pattern as described in claim 1 wherein said internal memory means contains a plurality of reference patterns that describe alphanumeric and punctuation symbols.

5. An apparatus for performing pattern recognition procedures regarding an unknown pattern as described in claim 4 wherein said pattern recognition procedures perform handwriting recognition and wherein said multiprocessing means is specially optimized for performing Dynamic Time Warping pattern recognition applications.

6. An apparatus for performing pattern recognition procedures regarding an unknown pattern as described in claim 4 wherein said pattern recognition procedures perform handwriting recognition and wherein said multiprocessing means is specially optimized for performing Hidden Markov Models pattern recognition applications.

7. An apparatus for performing pattern recognition procedures regarding an unknown pattern as described in claim 1 wherein

said two arithmetic pipelines, said two pointer pipelines and said control pipeline operate in parallel to execute Very Long Instruction Words within said pattern recognition procedures wherein said Very Long Instruction Words individually comprise said two arithmetic instructions and said two pointer instructions.

8. An apparatus for performing pattern recognition procedures regarding an unknown pattern as described in claim 7 wherein said arithmetic pipelines of said multiprocessing means are specially optimized to perform Dynamic Time Warping or Hidden Markov Models pattern recognition computations.

9. An apparatus for performing pattern recognition procedures regarding an unknown pattern as described in claim 7 wherein said pattern recognition procedures are handwriting recognition procedures, and wherein said computer system is a pen-based computer system.

10. A system for pattern recognition comprising:

general purpose host computer system means comprising;

a host bus means for providing communication within said general purpose host computer system means;

central processor means for executing host instructions, said central processor means coupled to said host bus means; and

host memory means for storing host information, said host memory means coupled to said host bus means; and

programmable multiprocessing means for executing pattern recognition instructions for comparing an unknown pattern against a plurality of reference patterns;

said programmable multiprocessing means coupled to said host bus means, wherein said programmable multiprocessing means further comprises;

(i) two arithmetic pipelines for executing two arithmetic instructions in parallel, a first arithmetic pipeline optimized for performing distance computations between segments of said unknown pattern and segments of said plurality of reference patterns and a second arithmetic pipeline optimized for performing best path computations in response to previously computed distance computations;

(ii) two pointer pipelines for executing two pointer instructions in parallel, each of said pointer pipelines for supplying pointer information for an associated arithmetic pipeline; and

(iii) a control pipeline for providing instruction flow control for said pattern recognition instructions, and wherein said two arithmetic pipelines, said two pointer pipelines and said control pipeline operate in parallel to execute Very Long Instruction Words within said pattern recognition instructions; and

private memory means coupled to said programmable multiprocessing means for storing a library of said plurality of reference patterns, said private memory means having at least one communication pathway to said programmable multiprocessing means that bypasses said host bus means wherein access of said private memory means by said programmable multiprocessing means does not necessarily increase information flow across said host bus means.

11. A system for pattern recognition as described in claim 10 further comprising:

stylus input means for interfacing with a user and for providing a communication signal representative of movement of a user hand; and

digitizer means for providing a digital signal representative of said communication signal, said digitizer means coupled to said host bus means and also coupled to said stylus input means to receive said communication signal.

12. A system for pattern recognition as described in claim 11 wherein said digitizer means supplies said unknown pattern representative of an unknown handwritten character for pattern recognition.

13. A system for pattern recognition as described in claim 10 wherein said central processor means of said host computer system means executes high level pattern recognition instructions that synchronize transfer of unknown patterns to said programmable multiprocessing means and that access results of said programmable multiprocessing means;

and

wherein said programmable multiprocessing means executes lower level pattern recognition instructions that compare said unknown pattern against a individual reference pattern of said plurality of reference patterns, said central processor means and said programmable multiprocessor means communicate over said host bus means to synchronize said high level and said lower level instructions.

14. A system for pattern recognition as described in claim 13 wherein said programmable multiprocessing means compares said unknown pattern against each of said reference patterns within said private memory means and reports matching information to said central processor means and wherein said central processor means determines, based on said matching information, reference patterns mostly closely matched to said unknown pattern.

15. A system for pattern recognition as described in claim 10 wherein said programmable multiprocessing means further comprises:

first internal buffer means for containing said unknown pattern and for containing at least one selected reference pattern of said plurality of reference patterns;

second internal buffer means for containing results generated from said programmable multiprocessing means including results of said distance and best path computations; and

program storage means for containing instructions executed by said programmable multiprocessing means.

16. A system for pattern recognition as described in claim 15 wherein said programmable multiprocessing means is optimized for Dynamic Time Warping or Hidden Markov Models pattern recognition procedures.

17. A system for pattern recognition as described in claim 10 wherein said programmable multiprocessing means is optimized for Dynamic Time Warping or Hidden Markov Models pattern recognition procedures.

18. In a computer system having a central processor for executing instructions, a host memory array for storing information, and a communication bus coupled to said central processor and to said host memory array, a pattern recognition apparatus coupled to said communication bus for determining an unknown pattern information, said pattern recognition apparatus comprising:

a programmable multiprocessor specially optimized for executing pattern recognition instructions, said programmable multiprocessor communicatively coupled with said central processor wherein said programmable multiprocessor further comprises;

(i) a first arithmetic pipeline for performing arithmetic instructions on individual points of said unknown pattern information against individual points of a selected reference pattern; and

(ii) a first pointer pipeline for executing pointer instructions to reference said individual points of said unknown pattern information and of said selected reference pattern, wherein said first pointer pipeline and said first arithmetic pipeline execute in parallel;

(iii) a second arithmetic pipeline for performing arithmetic instructions on results from said first arithmetic pipeline; and

(iv) a second pointer pipeline for executing pointer instructions to reference results from said first arithmetic pipeline, wherein said second pointer pipeline said second arithmetic pipeline, and said first arithmetic pipeline execute in parallel; and

a private memory array, separate from said host memory array, for storing reference patterns, said private memory array communicatively coupled to said programmable multiprocessor by at least one communication channel that bypasses said communication bus wherein communication between said private memory array and said programmable multiprocessor does not interfere with said control processor.

19. A pattern recognition apparatus as described in claim 18 wherein said pattern recognition instructions executed by said programmable multiprocessor are Very Long Instruction Words wherein said Very Long Instruction Words individually comprise two arithmetic instructions for execution by said first and second arithmetic pipelines and two pointer instructions for execution by said first and second pointer pipelines.

20. A pattern recognition apparatus as described in claim 18 wherein said programmable multiprocessor is specially optimized to perform handwriting recognition procedures using Dynamic Time Warping procedures.

21. A pattern recognition apparatus as described in claim 18 wherein said programmable multiprocessor is specially optimized to perform handwriting recognition procedures using Hidden Markov Models procedures.

22. A pattern recognition apparatus as described in claim 18 wherein said programmable multiprocessor is specially optimized to perform voice recognition procedures.

23. A pattern recognition apparatus as described in claim 18 wherein said programmable multiprocessor further comprises internal buffers for temporary storage of said unknown pattern information and for temporary storage of pattern recognition results between said unknown pattern information and said reference patterns.

24. A pattern recognition apparatus as described in claim 18 wherein said programmable multiprocessor is specially optimized to perform handwriting recognition procedures and further comprising:

handwriting stylus means for interfacing with a user to supply information regarding movement of a user hand; and

digitizer means coupled to said handwriting stylus for translating said information regarding movement of said user hand into digital form as an unknown pattern.

25. A pattern recognition apparatus as described in claim 18 wherein said programmable multiprocessor further comprises a first internal buffer for temporary storage of said unknown pattern information and said selected reference pattern, said first internal buffer referenced by pointers from said first pointer pipeline and wherein said programmable multiprocessor further comprises a second internal buffer for temporary storage of results from said first and said second arithmetic pipelines, said second internal buffer referenced by pointers from said second pointer pipeline.

26. A pattern recognition apparatus as described in claim 18 wherein said programmable multiprocessor performs handwriting pattern recognition and wherein said first arithmetic pipeline executes distance computations and wherein said second arithmetic pipeline executes best path computations.

27. In a computer system having a central processor for executing instructions, a host memory array for storing information, and a communication bus coupled to said central processor and to said host memory array, a pattern recognition apparatus for determining an unknown pattern, said pattern recognition apparatus coupled to said communication bus, said pattern recognition apparatus comprising:

a programmable multiprocessor specially optimized for executing pattern recognition instructions, said programmable multiprocessor comprising;

a first and a second arithmetic pipeline for executing two arithmetic instructions in parallel, said first arithmetic pipeline optimized for performing distance computations between segments of said unknown pattern and segments of a reference pattern of a plurality of reference patterns and said second arithmetic pipeline optimized for performing best path computations in response to previously computed distance computations;

a first and a second pointer pipeline for executing two pointer instructions in parallel, said first pointer pipeline supplying reference pointers to said first arithmetic pipeline and said second pointer pipeline supplying reference pointers to said second arithmetic pipeline;

a control pipeline for controlling instruction flow of said arithmetic and pointer instructions by executing a control instruction in parallel with said two arithmetic and pointer instructions; and

a private memory array, separate from said host memory array, for storing said plurality of reference patterns, said private memory array communicatively coupled to said programmable multiprocessor by at least one communication channel that bypasses said communication bus.

28. A pattern recognition apparatus as described in claim 27 and wherein said first and said second arithmetic pipelines, said first and said second pointer pipelines and said control pipeline operate in parallel to execute a Very Long Instruction Word comprising two arithmetic instructions and two pointer instructions.

29. A pattern recognition apparatus as described in claim 27 wherein said programmable multiprocessor is specially optimized for handwriting pattern recognition using Dynamic Time Warping procedures.

30. A pattern recognition apparatus as described in claim 27 wherein said programmable multiprocessor is specially optimized for handwriting pattern recognition using Hidden Markov Models procedures.

31. In a computer system having a central processor for executing instructions, a host memory array for storing information, and a communication bus coupled to said central processor and to said host memory array, an apparatus for performing pattern recognition instructions to reduce processing required of said central processor, said pattern recognition apparatus for determining an unknown pattern, said pattern recognition apparatus coupled to said communication bus, said pattern recognition apparatus comprising:

a) a programmable multiprocessor specially optimized for handwriting recognition and for executing pattern recognition instructions, said programmable multiprocessor comprising;

a first and a second arithmetic pipeline for executing two arithmetic instructions in parallel, said first arithmetic pipeline optimized for performing distance computations between segments of said unknown pattern and segments of a selected reference pattern and said second arithmetic pipeline optimized for performing best path computations in response to previously computed distance computations;

a first and a second pointer pipeline for executing two pointer instructions in parallel, said first pointer pipeline for supplying reference pointers to said first arithmetic pipeline and said second pointer pipeline for supplying reference pointers to said second arithmetic pipeline;

a control pipeline for controlling instruction flow of said arithmetic and pointer instructions by executing a control instruction in parallel with said two arithmetic and pointer instructions;

an internal buffer memory for storing said unknown pattern and said selected reference pattern; and

b) a private memory array, separate from said host memory array, for storing reference patterns, said private memory array communicatively coupled to said programmable multiprocessor by at least one communication channel that bypasses said communication bus.

32. A pattern recognition apparatus as described in claim 31 wherein said programmable multiprocessor further comprises memory to memory transfer means for transferring said selected reference pattern from said private memory array to said internal buffer memory over said communication channel that bypasses said communication bus.

33. A pattern recognition apparatus as described in claim 31 wherein said computer system is a pen-based computer system and wherein said computer system further includes a stylus for translating movement of a user hand and a digitizer coupled to said stylus for providing an unknown pattern to said programmable multiprocessor, said digitizer coupled to said communication bus.

34. A system for performing pattern recognition comprising:

host processor means for executing high level pattern recognition instructions and system tasks apart from said pattern recognition instructions;

communication bus means coupled to said host processor means, said communication bus means for supplying a communication pathway for said host processor means to perform said system tasks;

multiprocessor means specially optimized for performing lower level pattern recognition instructions between an unknown pattern and selected reference patterns, said multiprocessor means coupled to said communication bus means to receive instructions from said host processor means, wherein said multiprocessor means further comprises;

(i) a first arithmetic pipeline for performing arithmetic instructions on individual points of said unknown pattern against individual points of a selected reference pattern; and

(ii) a first pointer pipeline for executing pointer instructions to reference said individual points of said unknown pattern and of said selected reference pattern, wherein said first pointer pipeline and said first arithmetic pipeline execute in parallel;

(iii) a second arithmetic pipeline for performing arithmetic instructions on results from said first arithmetic pipeline; and

(iv) a second pointer pipeline for executing pointer instructions to reference said results from said first arithmetic pipeline, wherein said second pointer pipeline, said second arithmetic pipeline, and said first arithmetic pipeline execute in parallel; and

private memory means for storing a plurality of reference patterns, said private memory means coupled to said multiprocessor means by a communication pathway that bypasses said communication bus means wherein accessing of said selected reference patterns of said private memory means by said multiprocessor means does not increase information bandwidth over said communication bus means.

35. A system for performing pattern recognition as described in claim 34 wherein said multiprocessor means is specially optimized to perform handwriting recognition by Dynamic Time Warping or Hidden Markov Models procedures wherein said first arithmetic pipeline performs distance computations between said individual points of said unknown pattern against said individual points of said selected reference pattern and said second arithmetic pipeline performs best path computations on said results from said first arithmetic pipeline.

36. A system for performing pattern recognition as described in claim 35 further comprising:

handwriting stylus means for interfacing with a user to supply information regarding movement of a user hand; and

digitizer means coupled to said handwriting stylus for translating said information regarding movement of said user hand into digital form, said digitizer also for supplying an unknown pattern, said digitizer coupled to said communication bus means.

37. A system for performing pattern recognition as described in claim 35 wherein said high level pattern recognition instructions executed by said host processor means synchronize transferring of said unknown reference pattern into said multiprocessor means and also synchronize reporting of results of said multiprocessor means to said host processor means.

38. A user interactive system for pattern recognition comprising:

a pen-based host computer system comprising;

a host bus for providing communication within said pen-based host computer system;

central processor for executing high level pattern recognition instructions, said central processor coupled to said host bus;

an input stylus and digitizer responsive to movement of a user hand for receiving an unknown handwriting pattern, said input stylus and digitizer coupled to said host bus; and

a host memory array for storing host information, said host memory array coupled to said host bus; and

a programmable multiprocessor specially optimized for Dynamic Time Warping or Hidden Markov Models procedures for executing low level pattern recognition instructions to compare said unknown handwriting pattern against a plurality of reference handwriting patterns;

said programmable multiprocessor coupled to said host bus, wherein said programmable multiprocessor further comprises;

(i) a first arithmetic pipeline for performing arithmetic instructions on individual points of said unknown handwriting pattern information against individual points of a selected reference handwriting pattern to determine distance computations; and

(ii) a first pointer pipeline for executing pointer instructions to reference said individual points of said unknown handwriting pattern information and of said selected reference handwriting pattern, wherein said first pointer pipeline and said first arithmetic pipeline execute in parallel;

(iii) a second arithmetic pipeline for performing arithmetic instructions on results from said first arithmetic pipeline to determine best path computations; and

(iv) a second pointer pipeline for executing pointer instructions to reference said results from said first arithmetic pipeline, wherein said second pointer pipeline, said second arithmetic pipeline, and said first arithmetic pipeline execute in parallel; and

a private memory array coupled to said programmable multiprocessor for storing a library of said plurality of reference handwriting patterns, said private memory array having at least one communication pathway to said programmable multiprocessor that bypasses said host bus so that access of said private memory by said programmable multiprocessor does not necessarily increase information flow across said host bus.

39. A system for pattern recognition as described in claim 38 wherein said central processor controls transferring said unknown handwriting pattern to said programmable multiprocessor via said host bus and wherein said programmable multiprocessor controls transfer of selected reference handwriting patterns from said private memory to said programmable multiprocessor.

40. A computer implemented method of performing pattern recognition on an unknown pattern, said method comprising the computer implemented steps of:

a) executing high level pattern recognition instructions by a host processor;

b) executing lower level pattern recognition instructions by a programmable multiprocessor to compare said unknown pattern against selected reference patterns, said step of executing lower level pattern recognition instructions further comprising the steps of;

executing arithmetic instructions by a pair of arithmetic pipelines, said step of executing arithmetic instructions comprising the steps of performing arithmetic instructions on individual points of said unknown pattern against individual points of a selected reference pattern via a first arithmetic pipeline to perform distance computations and simultaneously performing arithmetic instructions on previous results from said first arithmetic pipeline via a second arithmetic pipeline to perform distance computations; and

executing pointer instructions by a pointer pipeline in parallel with said step of executing arithmetic instructions; and

c) communicating between said host processor and said programmable multiprocessor via a host computer bus; and

d) transferring said selected reference patterns from a private memory array to said programmable multiprocessor, said step of transferring bypassing said host computer bus.

41. A method of performing pattern recognition as described in claim 40 further comprising the computer implemented steps of:

transferring results of said programmable multiprocessor over said host computer bus to said host processor; and

identifying said unknown pattern based on said results of said programmable multiprocessor, said step of identifying performed by said host processor.

42. A method of performing pattern recognition as described in claim 41 wherein said method of pattern recognition is used for handwriting recognition and wherein said unknown pattern is representative of a handwritten character.

43. A method of performing pattern recognition as described in claim 40 wherein said high level pattern recognition instructions synchronize transfers of said unknown pattern to said programmable multiprocessor and also synchronize transfers of said results of said programmable multiprocessor to said host processor.

44. A method of performing pattern recognition as described in claim 43 wherein said lower level pattern recognition instructions perform point by point comparisons between said unknown pattern and said selected reference patterns to generate said results of said programmable multiprocessor.

46. A method of increasing accuracy and performance of pattern recognition applications as described in claim 43 wherein said step of performing high level pattern recognition procedures comprises the steps of:

receiving said unknown pattern via said input means;

transferring said unknown pattern to said programmable multiprocessing means;

transferring results of said programmable multiprocessing means to said central processor means; and

determining identity of said unknown pattern based on said results from said programmable multiprocessing means.

47. A method of increasing accuracy and performance of pattern recognition applications as described in claim 46 wherein said step of determining identity of said unknown pattern further comprises the step of displaying said identity to a display means.

45. In a computer system having a central processor means for executing computer system instructions, an input means for receiving an unknown pattern, a system memory array for storage of system information and a computer system bus for coupling said central processor means, said input means and said system memory array, a method of increasing accuracy and performance of pattern recognition applications comprising the computer implemented steps of:

performing high level pattern recognition procedures via said central processor means;

performing lower level pattern recognition procedures via a programmable multiprocessor means to reduce processing required of said central processor means, said step of executing lower level pattern recognition procedures further comprising the steps of;

(i) transferring a selected reference pattern from a private memory array to said programmable multiprocessor means;

(ii) comparing said selected reference pattern, point by point against said unknown pattern, said step of comparing comprising the steps of executing arithmetic instructions by a pair of arithmetic pipelines, said step of executing arithmetic instructions comprising the steps of performing arithmetic instructions on individual points of said unknown pattern information against individual points of a selected reference pattern via a first arithmetic pipeline to perform distance computations and simultaneously performing arithmetic instructions on previous results from said first arithmetic pipeline via a second arithmetic pipeline to perform best path computations and executing pointer instructions by a pointer pipeline in parallel with said step of executing arithmetic instructions; and

(iii) generating results of said comparing step indicating degrees of coincidence between said selected reference pattern and said unknown pattern;

communicating between said central processor means and said programmable multiprocessor means via said computer system bus;

storing said reference patterns in said private memory array; and

communicating between said private memory array and said programmable multiprocessor over a communication channel separate from said computer system bus to reduce bandwidth requirements of said computer system bus.

48. A method of increasing accuracy and performance of pattern recognition applications as described in claim 45 wherein said pattern recognition applications are handwriting recognition applications and wherein said programmable multiprocessor means is optimized to perform Dynamic Time Warping or Hidden Markov Models procedures.

49. In a computer system having a central processor means for executing computer system instructions, an input means for receiving an unknown information pattern, a system memory array for storage of system information and a computer system bus for coupling said central processor means, said input means and said system memory array, a method of increasing accuracy and performance of pattern recognition applications comprising the computer implemented steps of:

a) storing reference patterns in a private memory array;

b) performing high level pattern recognition procedures via said central processor means, said high level procedures comprising the steps of;

receiving an unknown pattern via said input means;

transferring said unknown pattern to a programmable multiprocessing means via said computer system bus;

transferring results of said programmable multiprocessing means to said central processor means via said computer system bus; and

determining identity of said unknown pattern based on said results from said programmable multiprocessing means; and

c) performing lower level pattern recognition procedures via said programmable multiprocessor means to reduce processing required of said central processor means, said lower level pattern recognition procedures comprises the steps of;

transferring a selected reference pattern from said private memory array to said programmable multiprocessor means via a communication channel separate from said computer system bus to reduce bandwidth requirements of said computer system bus;

comparing said selected reference pattern, point by point, against said unknown pattern, said step of comparing comprising the steps of executing arithmetic instructions by a pair of arithmetic pipelines, said step of executing arithmetic instructions comprising the steps of performing arithmetic instructions on individual points of said unknown pattern information against individual points of a selected reference pattern via a first arithmetic pipeline to perform distance computations and simultaneously performing arithmetic instructions on previous results from said first arithmetic pipeline via a second arithmetic pipeline to perform best path computations and executing pointer instructions by a pointer pipeline in parallel with said step of executing arithmetic instructions; and

generating results of said comparing step indicating degrees of coincidence between said selected reference pattern and said unknown pattern.

50. A method of increasing accuracy and performance of pattern recognition applications as described in claim 49 wherein said programmable multiprocessor is specially optimized for Dynamic Time Warping or Hidden Markov Models pattern recognition and wherein said pattern recognition application is a handwriting recognition application.