ROBUST PERSONALIZATION THROUGH BIASED REGULARIZATION

• US 20070239450A1
• Filed: 04/06/2006
• Published: 10/11/2007
• Est. Priority Date: 04/06/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Chellapilla, Kumar, Kienzle, Wolf
• Granted Patent: US 7,886,266 B2
• Time in Patent Office: Days
• US Class Current: 704/244
• CPC Class Codes: G10L 15/07 to the speaker

Abstract

The subject disclosure pertains to systems and methods for personalization of a recognizer. In general, recognizers can be used to classify input data. During personalization, a recognizer is provided with samples specific to a user, entity or format to improve performance for the specific user, entity or format. Biased regularization can be utilized during personalization to maintain recognizer performance for non-user specific input. In one aspect, regularization can be biased to the original parameters of the recognizer, such that the recognizer is not modified excessively during personalization.

Claims

1. A system for enhancing performance of a data pattern recognizer for an entity, comprising:

a personalization component that updates the recognizer based at least in part upon at least one sample customized to the entity, the personalization component utilizes regularization favoring a base set of parameters of the recognizer during update of the recognizer.

2. The system of claim 1, the personalization component further comprises:

a base parameter component that obtains the base parameters of the recognizer;

a regularization component that utilizes regularization to determine a new set of parameters for the recognizer; and

a parameter update component that updates the base set of parameters of the recognizer to the new set of parameters.

3. The system of claim 2, the regularization component further comprises:

a deviation component that determines the deviation from the base set of parameters to a possible set of parameters;

an error rate component that determines the error rate over the at least one customized sample using the possible set of parameters; and

a minimization component that selects the new set of parameters to minimize the sum of the error rate and the deviation.

4. The system of claim 2, the base parameter component maintains the base set of parameters of the recognizer.

5. The system of claim 2, the base parameter component obtains the base set of parameters of the recognizer from the recognizer.

6. The system of claim 1, the recognizer classifies image data, the image data includes at least one of printed text and handwritten text.

7. The system of claim 1, further comprising:

a sample receiving component that receives the at least one customized sample, the at least one customized sample includes raw data and a corresponding label.

8. The system of claim 7, the personalization component adds the corresponding label of the customized sample to a set of recognized labels of the recognizer.

9. The system of claim 1, the regularization utilized is Tichonov regularization.

15. The method of claim 1, further comprising maintaining the base state of the trained recognizer.

16. The method of claim 1, further comprising obtaining the base state of the trained recognizer from the trained recognizer.

10. A method for enhancing the performance of a trained recognizer for an individual, comprising:

retraining the trained recognizer based at least in part upon at least one sample personalized to the individual using regularization biased to a base state of the trained recognizer.

11. The method of claim 10, retraining the trained recognizer, further comprises:

determining a new state of the trained recognizer based upon the regularization and the at least one personalized sample; and

updating the trained recognizer.

12. The method of claim 11, determining the new state of the trained recognizer further comprises:

determining a distance between the base state of the trained recognizer and a possible state of the trained recognizer;

determining risk for the at least one personalized sample using the possible state of the trained recognizer; and

minimizing the distance and the risk.

13. The method of claim 10, the at least one personalized sample includes image data, the image data includes at least one of printed text and handwritten text.

14. The method of claim 10, retraining the trained recognizer is based at least in part on a weight for the at least one personalized sample, the weight based in part on the likelihood of recognition error for the personalized sample.

17. A system for enhancing data classification for a recognizer for an individual, comprising:

means for training the recognizer based at least in part upon at least one sample generated by the individual using regularization biased to a base weights of the recognizer.

18. The system of claim 17, further comprising:

means for selecting a new set of weights for the recognizer based upon the regularization and the at least one individual sample; and

means for updating the base weights of the recognizer.

19. The system of claim 18, means for selecting the set of weights further comprises:

means for determining the deviation between the set of base weights of the recognizer and a possible set of weights; and

means for determining an error rate for the at least one individual sample using the possible set of weights, selection of the new set of weights minimizes the deviation and the error rate.

20. The system of claim 17, the recognizer classifies image data, image data includes at least one of printed text and handwritten text.