Elastic distortions for automatic generation of labeled data

• US 20050025355A1
• Filed: 07/31/2003
• Published: 02/03/2005
• Est. Priority Date: 07/31/2003
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Simard, Patrice Y., Steinkraus, David W.
• Granted Patent: US 7,418,128 B2
• Time in Patent Office: Days
• US Class Current: 382/159
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06N 20/20 Ensemble learning
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

A system that facilitates generation of data that can be employed in connection with training a classifier. The system comprises a component that receives a data set that is employed in connection with training the classifier, and an expansion component that applies elastic distortion algorithm(s) to a subset of the data set to generate additional labeled training data.

Claims

1. A system that facilitates training a classifier, comprising:

a component that receives a data set to be employed in connection with training the classifier; and

an expansion component that applies elastic distortion algorithm(s) to a subset of the data set to generate additional training data.

2. A classifier trained using the additional data generated via the system of claim 1.

3. The classifier of claim 2 employed for at least one of handwriting, speech, and object analysis.

4. The classifier of claim 2 being at least one of a neural network, a Bayesian belief network, a support vector machine, a boosting algorithm, and a variational method system.

5. The classifier of claim 2 being at least one of a convolutional neural network and a two layer Multi-Layer Perceptron neural network.

6. The system of claim 1, wherein the system can infer transformation invariance of the data set.

7. The system of claim 1, the received data set being digital image(s) defined by one or more pixel(s).

8. The system of claim 7, the expansion component associated with a random number generator employed to facilitate creation of a random displacement field via displacing the pixel(s) according to output values of the random number generator.

9. The system of claim 8, further comprising a low-pass filter that is applied to the displacement field to facilitate smoothing such displacement field.

10. The system of claim 7, wherein the elastic distortions of one or more digital image(s) are obtained by applying a smooth displacement field to the digital image(s).

11. The system of claim 10, wherein a random number generator is employed to facilitate generation of the smooth displacement field.

12. The system of claim 10, further comprising:

a low order polynomial; and

a random number generator, wherein values output by the random number generator are utilized as coefficient(s) of the low order polynomial.

13. The system of claim 10, wherein the smooth displacement field is obtained by over-sampling a low resolution displacement field, the low resolution displacement field generated at least in part by random numbers output from a random number generator.

14. The system of claim 10 wherein the smooth displacement field is obtained by applying a low-pass filter over a noisy field.

15. The system of claim 14, wherein convolution techniques are utilized as the low-pass filter.

16. The system of claim 15, wherein the displacement field is convolved with a Gaussian of standard deviation σ

.

17. The system of claim 16, the displacement field convolved in one dimension at a time.

18. The system of claim 9, further comprising an artificial intelligence component that infers a type of low-pass filter and parameters of the filter to be utilized given at least one of user, user state, user history, computing state, available computing resources, and classifier architecture.

19. The system of claim 9, further comprising a scaling component for scaling the filtered displacement field.

20. The system of claim 9, the displacement field applied to the image(s) to elastically distort such image(s).

21. The system of claim 20, further comprising an interpolation component that facilitates computation of position(s) of pixel(s) that define the elastically distorted image(s) and appropriate color value(s) relating to such position(s).

22. A classifier trained at least in part upon the elastically distorted image(s) of claim 21, the classifier employed to at least one of recognize and analyze one or more of handwriting, speech, and object(s).

23. The system of claim 1, a subset of the received data set describes a digital trajectory(ies), wherein the digital trajectory(ies) are defined by point(s) that have associated position and time values to enable the trajectory(ies) to be defined by equations x(t) and y(t).

24. The system of claim 23, the subset of the received data set obtained via a pressure-sensitive screen, and the time-values obtained via time-sampling inputs to the pressure-sensitive screen.

25. The system of claim 23, the time value(s) obtained via assigning time values to known positions according to at least one of known human writing, speaking, and drawing techniques.

26. The system of claim 23, the expansion component associated with a random number generator employed to facilitate creation of a random displacement field as a function of time, the displacement field being added to the equations x(t) and y(t).

27. The system of claim 23, wherein elastic distortions of a trajectory are obtained by adding a smooth displacement field to the trajectory.

28. The system of claim 27, wherein a random number generator is employed to facilitate generation of the smooth displacement field.

29. The system of claim 27, further comprising:

a low order polynomial; and

a random number generator, wherein values output by the random number generator are utilized as coefficient(s) of the low order polynomial.

30. The system of claim 27, wherein the smooth displacement field is obtained by over-sampling a low resolution displacement field, the low resolution displacement field generated at least in part by random numbers output from a random number generator.

31. The system of claim 27 wherein the smooth displacement field is obtained by applying a low pass filter over a noisy field.

32. The system of claim 31, wherein convolution techniques are utilized as the low-pass filter.

33. The system of claim 31, further comprising an artificial intelligence component that infers a type of low-pass filter and parameters of the filter to be utilized given at least one of user, user state, user history, computing state, available computing resources, and classifier architecture.

34. The system of claim 31, wherein convolution techniques are utilized as the low-pass filter.

35. The system of claim 34, wherein the displacement field is convolved with a Gaussian of standard deviation σ

.

36. The system of claim 31, further comprising a scaling component that scales the filtered displacement field.

37. The system of claim 36, wherein x(t) and y(t) are re-generated with time value(s) according to the smoothed displacement field to thereby elastically distort the image.

38. A classifier trained at least in part upon the elastically distorted trajectory(ies) of claim 37, the classifier employed to at least one of recognize and analyze one or more of handwriting, speech, and object(s).

68. A computer readable medium that has stored thereon the computer executable components of claim 1.

72. A portable computing device that employs the system of claim 11.

39. A method for generating data that can be utilized in connection with training a classifier employed to analyze at least one of handwriting, speech, and or objects, comprising:

receiving a data set formatted as one or more image(s) with pixel(s) that define the one or more image(s);

generating a smooth displacement field using randomly generated numbers; and

elastically distorting the image(s) based at least in part upon the displacement field.

40. The method of claim 39, further comprising interpolating pixel(s) to compute the elastically distorted images.

41. The method of claim 39, wherein convolution techniques are employed to generate the smooth displacement field.

42. The method of claim 41, further comprising applying a Gaussian of standard deviation σ

to the displacement field to facilitate generation of the smooth displacement field.

43. The method of claim 39, wherein a polynomial is employed to generate the displacement field.

44. The method of claim 39, further comprising inferring transformation invariance of the one or more image(s).

45. The method of claim 39, further comprising multiplying the smooth displacement field by a scalar.

46. The method of claim 45, further comprising relaying the elastically distorted image(s) to a classifier.

50. The method of claim 46, wherein a polynomial is employed to filter the displacement field.

51. The method of claim 46, further comprising inferring transformation invariance of the one or more image(s).

52. The method of claim 46, further comprising multiplying the filtered displacement field by a scalar.

53. The method of claim 46, further comprising relaying the elastically distorted image(s) to a classifier.

66. The system of claim 59, further comprising:

means for applying time value(s) of the displacement field of claim 50 to the functions x(t) and y(t).

69. A computer readable medium that has stored thereon computer executable instructions for performing the method of claim 39.

47. A method for generating data that can be utilized in connection with training a classifier employed to analyze at least one of handwriting, speech, and or objects, comprising:

receiving a data set formatted as one or more trajectory(ies) that define one or more pattern(s) to classify;

associating time value(s) with coordinates of point(s) associated with the trajectory(ies), thereby enabling description of the one or more trajectory(ies) by functions x(t) and y(t);

generating a smooth displacement field via using randomly generated numbers;

adding the smooth displacement field to the functions x(t) and y(t) to generate new elastically distorted trajectories x(t) and y(t).

48. The method of claim 47, wherein convolution techniques are employed as a low-pass filter to generate the smooth displacement.

49. The method of claim 48, further comprising applying a Gaussian of standard deviation σ

to the displacement field to filter such displacement field.

70. A computer readable medium that has stored thereon computer executable instructions for performing the method of claim 47.

73. A portable computing device that employs the method of claim 47.

54. A system that facilitates generation of labeled data employed in connection with training a classifier, comprising:

means for receiving a data set formatted as one or more image(s), the one or more image(s) defined by one or more pixel(s); and

means for elastically distorting the one or more image(s).

55. The system of claim 54, further comprising:

means for generating random value(s) pertaining to coordinate(s) of the pixel(s); and

means for generating a displacement field based at least in part upon the random value(s).

56. The system of claim 55, further comprising:

means for smoothing the displacement field; and

means for scaling the displacement field.

71. A computer readable medium that has stored thereon the computer executable system of claim 54.

74. A portable computing device employing the system of claim 54.

57. A system that facilitates generation of labeled data employed in connection with training a classifier, comprising:

means for receiving a data set formatted as one or more trajectories, the one or more trajectory(ies) defined by one or more point(s); and

means for elastically distorting the one or more trajectory(ies).

58. The system of claim 57, further comprising:

means for associating time value(s) to the point(s), thereby enabling defining of the one or more trajectory(ies) by functions x(t) and y(t);

means for generating random value(s) pertaining to the time value(s); and

means for creating a smooth displacement field based at least in part upon the random value(s).

59. The system of claim 58, further comprising:

means for filtering the displacement field to filter high-frequency displacements; and

means for scaling the displacement field.

60. The system of claim 58, the means for creating a smooth displacement field comprising:

a low order polynomial; and

a random number generator, wherein values output by the random number generator are utilized as coefficient(s) of the low order polynomial.

61. The system of claim 58, the means for creating a smooth displacement field comprising over-sampling a low resolution displacement field, the low resolution displacement field generated at least in part by random numbers output from a random number generator.

62. The system of claim 58, the means for creating a smooth displacement field comprising applying a low-pass filter over a noisy field.

63. The system of claim 62, wherein convolution techniques are utilized as the low-pass filter.

64. The system of claim 63, wherein the displacement field is convolved with a Gaussian of standard deviation σ

.

65. The system of claim 64, the displacement field convolved in one dimension at a time.

67. A data packet that passes between at least two computer processes, comprising:

a field that stores an elastically distorted image, the image distorted via application of a smooth displacement field to an initial image, the displacement field generated at least in part by random values output from a random number generator.

75. A portable computing device, comprising:

a component that receives a human generated input; and

a classifier that was trained via employment of an expansion component that applies elastic distortion algorithm(s) to a subset of a training data set to generate additional training data.