Active machine learning

• US 10,262,272 B2
• Filed: 12/07/2014
• Issued: 04/16/2019
• Est. Priority Date: 12/07/2014
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Inventors: Chickering, David Maxwell, Meek, Christopher A., Simard, Patrice Y., Iyer, Rishabh Krishnan
• Primary Examiner(s): Chaki, Kakali
• Assistant Examiner(s): Smith, Brian M
• Application Number:
  US14/562,747
• Publication Number:
  US 20160162802A1
• Time in Patent Office: 1,591 Days
• Field of Search: None
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

Technologies are described herein for active machine learning. An active machine learning method can include initiating active machine learning through an active machine learning system configured to train an auxiliary machine learning model to produce at least one new labeled observation, refining a capacity of a target machine learning model based on the active machine learning, and retraining the auxiliary machine learning model with the at least one new labeled observation subsequent to refining the capacity of the target machine learning model. Additionally, the target machine learning model is a limited-capacity machine learning model according to the description provided herein.

Claims

1. A method comprising:

initiating active machine learning through an active machine learning system configured to train an auxiliary machine learning model;

evaluating an unlabeled observation using the auxiliary machine learning model to generate a first score;

evaluating the unlabeled observation using a target machine learning model to generate a second score;

comparing the first score to the second score to calculate a magnitude of a difference between the first score and the second score;

identifying a machine learning feature using the magnitude;

updating the target machine learning model based at least on a refinement using the machine learning feature, wherein the target machine learning model includes a limited-capacity machine learning model;

retraining the auxiliary machine learning model with at least one new labeled observation subsequent to updating the target machine learning model, wherein the retrained version of the auxiliary machine learning model produces the at least one new labeled observation using the machine learning feature from the unlabeled observation subsequent to refining the capacity of the target machine learning model; and

providing the updated target machine learning model to a computing device capable of computation of device labels using the updated target machine learning model.

2. A method as claim 1 recites, wherein the auxiliary machine learning model includes a capacity larger than the target machine learning model.

3. A method as claim 1 recites, wherein the auxiliary machine learning model includes a semantic machine learning model.

4. A method as claim 3 recites, wherein the semantic machine learning model includes a bag-of-words machine learning model.

5. A method as claim 1 recites, wherein initiating the active machine learning comprises:

selecting one or more unlabeled observations from a pool of unlabeled observations.

6. A method as claim 5 recites, wherein the refining the capacity includes incrementally adding or removing features from the target machine learning model based on an output of the auxiliary machine learning model responsive to processing the one or more unlabeled observations.

7. A method as claim 1 recites, wherein the refining the capacity includes incrementally adding or removing features from the target machine learning model based at least on the initiated active machine learning of the auxiliary machine learning model.

8. A method as claim 1 recites, further comprising implementing diversity in the initiated active machine learning by at least one submodular function.

9. A method as claim 1 recites, further comprising implementing diversity in the initiated active machine learning by establishing a subset labelset in a pool of unlabeled observations configured to provide diverse unlabeled observations from the subset labelset.

10. A method as claim 9 recites, further comprising selecting one or more unlabeled observations from the subset labelset for processing by the auxiliary machine learning model.

11. A method as claim 1 recites, further comprising reducing a color-blindness of the target machine learning model based at least on disagreement between the auxiliary machine learning model and the target machine learning model.

12. A non-transitory computer-readable medium having computer-executable instructions thereupon that, when executed by a computer, cause the computer to perform operations comprising:

selecting an unlabeled observation from a pool of unlabeled observations through an auxiliary machine learning model, wherein it is not known to which one of a plurality of classes the unlabeled observation belongs;

evaluating the unlabeled observation using an auxiliary machine learning model to generate a first score;

evaluating the unlabeled observation using a target machine learning model to generate a second score;

comparing the first score to the second score to calculate a magnitude of a difference between the first score and the second score;

identifying a machine learning feature using the magnitude;

updating the target machine learning model based on a refinement using the machine learning feature, wherein the target machine learning model is a limited-capacity machine learning model;

retraining the auxiliary machine learning model with at least one new labeled observation subsequent to updating the target machine learning model, wherein the retrained version of the auxiliary machine learning model produces the at least one new labeled observation using the machine learning feature from the unlabeled observation subsequent to refining the capacity of the target machine learning model; and

providing the updated target machine learning model to a computing device capable of computation of device labels using the updated target machine learning model.

13. A computer-readable medium as claim 12 recites, wherein the refining the capacity includes:

incrementally adding at least one feature to the target machine learning model based on features contained within the new labeled observation; and

incrementally removing at least one feature from the target machine learning model based on the features contained within the new labeled observation.

14. A computer-readable medium as claim 12 recites, wherein the selecting the unlabeled observation includes selecting the unlabeled observations based at least on optimization of at least one submodular function.

15. A computer-readable medium as claim 12 recites, wherein the selecting the unlabeled observation includes selecting the unlabeled observations from a subset labelset in the pool of unlabeled observations, the subset labelset configured to provide diverse unlabeled observations.

16. A computer-readable medium as claim 12 recites, wherein the refining the capacity of the target machine learning model includes reducing a color-blindness of the target machine learning model based at least on disagreement between the auxiliary machine learning model and the target machine learning model.

17. An active machine learning system, the system comprising:

at least one processor; and

memory including instructions that, when executed by the at least one processor, cause the at least one processor to perform operations to;

assign a first score to an unlabeled observation using an auxiliary machine learning model;

assign a second score to the unlabeled observation using a target machine learning model, wherein the target machine learning model and the auxiliary machine learning model are from different machine learning model classes, and wherein the target machine learning model is a limited-capacity machine learning model;

compare the first score and the second score to calculate a magnitude of a difference between the first score and the second score;

identify a machine learning feature using the magnitude;

update the target machine learning model based at least on a refinement using the machine learning feature;

retrain the auxiliary machine learning model with at least one new labeled observation subsequent to the update of the target machine learning model, wherein the retrained version of the auxiliary machine learning model produces the at least one new labeled observation using the machine learning feature from the unlabeled observation subsequent to refining the capacity of the target machine learning model; and

provide the updated target machine learning model to a computing device capable of computation of device labels using the updated target machine learning model.

18. A system as claim 17 recites, wherein the instructions to compare the first score and the second score further comprises instructions to:

determining that the target machine learning model has returned a false positive when the magnitude is negative; and

determining that the target machine learning model has returned a false negative when the magnitude is positive.

19. A system as claim 18 recites, further comprising instructions to:

extend a scope of the target machine learning model to include a new feature previously not within the scope of the target machine learning model when the target machine learning model has returned a false positive.

20. A system as claim 18 recites, further comprising instructions to:

narrow a scope of the target machine learning model to remove a feature previously within the scope of the target machine learning model when the target machine learning model has returned a false positive.