Method for improving performance of a trained machine learning model

1. A computer-implemented method for improving performance of a trained machine learning model, comprising:

receiving, at a second classifier with a second objective function, a first classification output from a first classifier with a first objective function, the first classification output generated by classifying an object identified in an input received at the first classifier; and

classifying, at the second classifier, the object according to features generated by convolving the first classification output with weights of the second classifier, the second classifier directly reducing a number of classification errors of the first classification output.

2. The computer-implemented method of claim 1 , in which the first objective function is differentiable.

3. The computer-implemented method of claim 1 , in which the second objective function is non-differentiable.

4. The computer-implemented method of claim 1 , in which the second objective function is a function of a difference between errors of the first classifier and the second classifier.

5. The computer-implemented method of claim 1 , further comprising determining the second objective function based at least in part on a mixture of probabilities from a higher complexity model.

6. The computer-implemented method of claim 1 , further comprising adding the second classifier without retraining the first classifier.

7. The computer-implemented method of claim 1 , further comprising adding the second classifier externally to the first classifier.

8. The computer-implemented method of claim 1 , further comprising assigning weights to features produced by a model, which is trained by the first classifier, to an identity value.

9. The computer-implemented method of claim 8 , further comprising assigning weights to features produced by a probability vector of a high complexity model to zero.

10. The computer-implemented method of claim 1 , further comprising assigning weights to features produced by a probability vector of the second classifier.

11. The computer-implemented method of claim 1 , further comprising assigning weights to features produced by a probability vector of a high complexity model to zero.

12. The computer-implemented method of claim 11 , further comprising assigning weights to features produced by a probability vector of the second classifier.

13. The computer-implemented method of claim 1 , farther comprising scaling probability vectors generated by a higher complexity model by a fixed temperature T.

14. An apparatus for improving performance of a trained machine learning model, comprising:

a memory; and

at least one processor coupled to the memory, the at least one processor configured:

to receive, at a second classifier with a second objective function, a first classification output from a first classifier with a first objective function, the first classification output generated by classifying an object identified in an input received at the first classifier; and

to classify, at the second classifier, the object according to features generated by convolving the first classification output with weights of the second classifier, the second classifier directly reducing a number of classification errors of the first classification output.

15. The apparatus of claim 14 , in which the first objective function is differentiable.

16. The apparatus of claim 14 , in which the second objective function is non-differentiable.

17. The apparatus of claim 14 , in which the second objective function is a function of a difference between errors of the first classifier and the second classifier.

18. The apparatus of claim 14 , in which the at least one processor is further configured to determine the second objective function based at least in part on a mixture of probabilities from a higher complexity model.

19. The apparatus of claim 14 , in which the at least one processor is further configured to add the second classifier without retraining the first classifier.

20. The apparatus of claim 14 , in which the at least one processor is further configured to add the second classifier externally to the first classifier.

21. The apparatus of claim 14 , in which the at least one processor is further configured to assign weights to features produced by a model, which is trained by the first classifier, to an identity value.

22. The apparatus of claim 21 , in which the at least one processor is further configured to assign weights to features produced by a probability vector of a high complexity model to zero.

23. The apparatus of claim 14 , in which the at least one processor is further configured to assign weights to features produced by a probability vector of the second classifier.

24. The apparatus of claim 14 , in which the at least one processor is further configured to assign weights to features produced by a probability vector of a high complexity model to zero.

25. The apparatus of claim 24 , in which the at least one processor is further configured to assign weights to features produced by a probability vector of the second classifier.

26. The apparatus of claim 14 , in which the at least one processor is further configured to scale probability vectors generated by a higher complexity model by a fixed temperature T.

27. An apparatus for improving performance of a trained machine learning model, comprising:

means for receiving, at a second classifier with a second objective function, a first classification output from a first classifier with a first objective function, the first classification output generated by classifying an object identified in an input received at the first classifier; and

means for classifying, at the second classifier, the object according to features generated by convolving the first classification output with weights of the second classifier, the second classifier directly reducing a number of classification errors of the first classification output.

28. A non-transitory computer-readable medium having encoded thereon program code for improving performance of a trained machine learning model, the program code being executed by a processor and comprising:

program code to receive, at a second classifier with a second objective functions, a first classification output from a first classifier with a first objective function, the first classification output generated by classifying an object identified in an input received at the first classifier; and

program code to classify, at the second classifier, the object according to features generated by convolving the first classification output with weights of the second classifier, the second classifier directly reducing a number of classification errors of the first classification output.