Loss augmentation for predictive modeling
A machine learning system that incorporates arbitrary constraints into deep learning model is provided. The machine learning system provides a set of penalty data points en a set of arbitrary constraints in addition to a set of original training data points. The machine learning system assigns a penalty to each penalty data point in the set of penalty data points. The machine learning system optimizes a machine learning model by solving an objective function based on an original loss function and a penalty loss function. The original loss function is evaluated over a set of original training data points and the penalty loss function is evaluated over the set of penalty data points. The machine learning system provides the optimized machine learning model based on a solution of the objective function.
1. A computing device comprising:
a processor; and
a storage device storing a set of instructions, wherein an execution of the set of instructions by the processor configures the computing device to perform acts comprising:
providing a set of penalty data points enforcing a set of arbitrary constraints in addition to a set of original training data points for a deep learning model;
assigning a penalty to each penalty data point in the set of penalty data points;
optimizing the deep learning model by solving an objective function based on an original loss function and a penalty loss function, wherein the original loss function is evaluated over a set of original training data points and the penalty loss function is evaluated over the set of penalty data points; and
providing the optimized deep learning model based on a solution of the objective function.
2. The computing device of claim 1 , wherein optimizing the deep learning model further comprises maximizing the penalty loss function over additional data points in addition to the set of penalty data points.
3. The computing device of claim 2 , wherein the set of instructions further configures the computing device to perform acts comprising: upon determining that the maximized penalty loss function causes the objective function to be greater than a threshold, updating the set of penalty data points with the additional data points and continuing to optimize the deep learning model by solving the objective function.
4. The computing device of claim 2 , wherein the set of instructions further configures the computing device to perform an act comprising: upon determining that the maximized penalty loss function causes the objective function to be less than a threshold, terminating optimization of the deep learning model.
5. The computing device of claim 2 , wherein the additional data points are samples identified based on an earlier iteration of a stochastic gradient descent operation used to optimize the deep learning.
6. The computing device of claim 5 , wherein the additional data points are identified based on a violation of the arbitrary constraints during the earlier iteration of the stochastic gradient descent operation.
7. The computing device of claim 1 , wherein the original loss function with the penalty loss function are additive terms in the objective function.
8. The computing device of claim 1 , wherein the deep learning model comprises one or more intermediate layers.
9. A computer-implemented method comprising:
providing a set of penalty data points enforcing a set of arbitrary constraints in addition to a set of original training data points for a deep learning model;
assigning a penalty to each penalty data point in the set of penalty data points;
optimizing the deep learning model by solving an objective function based on an original loss function and a penalty loss function, wherein the original loss function is evaluated over a set of original training data points and the penalty loss function is evaluated over the set of penalty data points; and
providing the optimized deep learning model based on a solution of the objective function.
10. The computer-implemented method of claim 9 , wherein optimizing the deep learning model further comprises maximizing the penalty loss function over additional data points in addition to the set of penalty data points.
11. The computer-implemented method of claim 10 , further comprising, upon determining that the maximized penalty loss function causes the objective function to be greater than a threshold, updating the set of penalty data points with the additional data points and continuing to optimize the deep learning model by solving the objective function.
12. The computer-implemented method of claim 10 , further comprising, upon determining that the maximized penalty loss function causes the objective function to be less than a threshold, terminating optimization of the deep learning model.
13. The computer-implemented method of claim 10 , wherein the additional data points are samples identified based on an earlier iteration of a stochastic gradient descent operation used to optimize the deep learning.
14. The computer-implemented method of claim 13 , wherein the additional data points are identified based on a violation of the arbitrary constraints during the earlier iteration of the stochastic gradient descent operation.
15. The computer-implemented method of claim 9 , wherein the original loss function with the penalty loss function are additive terms in the objective function.
16. The computer-implemented method of claim 9 , wherein the deep learning model comprises one or more intermediate layers.
17. A computer program product comprising:
one or more non-transitory computer-readable storage devices and program instructions stored on at least one of the one or more non-transitory storage devices, the program instructions executable by a processor, the program instructions comprising sets of instructions for:
providing a set of penalty data points enforcing a set of arbitrary constraints in addition to a set of original training data points for a deep learning model;
assigning a penalty to each penalty data point in the set of penalty data points;
optimizing the deep learning model by solving an objective function based on an original loss function and a penalty loss function, wherein the original loss function is evaluated over a set of original training data points and the penalty loss function is evaluated over the set of penalty data points; and
providing the optimized deep learning model based on a solution of the objective function.
18. The computer program product of claim 17 , wherein optimizing the deep learning model further comprises maximizing the penalty loss function over additional data points in addition to the set of penalty data points.
19. The computer program product of claim 17 , wherein the original loss function with the penalty loss function are additive terms in the objective function.
20. The computer program product of claim 17 , wherein the deep learning model comprises one or more intermediate layers.