Reinforcement learning based locally interpretable models

1. A method for training a locally interpretable model, the method comprising:

obtaining, at data processing hardware, a set of training samples;

training, by the data processing hardware, using the set of training samples, a black-box model;

generating, by the data processing hardware, using the trained black-box model and the set of training samples, a set of auxiliary training samples;

training, by the data processing hardware, using the set of auxiliary training samples, a baseline interpretable model;

training, by the data processing hardware, using the set of auxiliary training samples and the baseline interpretable model, an instance-wise weight estimator model;

for each auxiliary training sample in the set of auxiliary training samples, determining, by the data processing hardware, using the trained instance-wise weight estimator model, a selection probability for the auxiliary training sample;

selecting, by the data processing hardware, based on the selection probabilities, a subset of auxiliary training samples from the set of auxiliary training samples; and

training, by the data processing hardware, using the subset of auxiliary training samples, the locally interpretable model.

2. The method of claim 1 , wherein training the instance-wise weight estimator model comprises:

determining a reinforcement signal; and

updating the instance-wise weight estimator model based on the reinforcement signal.

3. The method of claim 2 , wherein determining the reinforcement signal comprises:

determining first loss data for the locally interpretable model based on a loss function;

determining second loss data for the baseline interpretable model based on the loss function; and

determining the reinforcement signal based on the first loss data for the locally interpretable model and the second loss data for the baseline interpretable model.

4. The method of claim 1 , wherein selecting the subset of auxiliary training samples from the set of auxiliary training samples comprises, for each auxiliary training sample in the set of auxiliary training samples:

determining, based on the selection probability, a corresponding selection value indicating either selection or no selection; and

when the corresponding selection value indicates selection, adding the auxiliary training sample into the subset of auxiliary training samples.

5. The method of claim 4 , wherein determining, for each auxiliary training sample, the selection probability for the auxiliary training sample comprises:

receiving, at the instance-wise weight estimator model, a validation sample; and

determining, by the instance-wise weight estimator model, the selection probability based on the validation sample.

6. The method of claim 1 , further comprising, after training the instance-wise weight estimator model:

receiving, at the instance-wise weight estimator model, a validation sample;

generating, by the instance-wise weight estimator model, the selection probability for each auxiliary training sample in the set of auxiliary training samples; and

training the locally interpretable model based on the validation sample and the selection probabilities of each auxiliary training sample in the set of auxiliary training samples.

7. The method of claim 1 , wherein training the black-box model comprises minimizing a mean squared error loss function or a cross-entropy loss function.

8. The method of claim 1 , wherein each training sample in the set of training samples comprises training data and a given label.

9. The method of claim 8 , wherein generating the set of auxiliary training samples comprises, for each training sample in the set of training samples:

obtaining, from the trained black-box model, using the training data of the training sample, a predicted label for the training data;

combining the training data and the predicted label into an auxiliary training sample; and

adding the auxiliary training sample to the set of auxiliary training samples.

10. The method of claim 1 , wherein training the instance-wise weight estimator model comprises:

obtaining a set of probe training samples;

for each probe training sample in the set of probe training samples, determining a selection probability;

determining, based on the selection probability of each probe training sample, an instance-wise selection vector;

optimizing the locally interpretable model based on the instance-wise selection vector; and

updating the instance-wise weight estimator model based on loss data from the optimized locally interpretable model.

11. A system comprising:

data processing hardware; and

memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations comprising:

obtaining a set of training samples;

training, using the set of training samples, a black-box model;

generating, using the trained black-box model and the set of training samples, a set of auxiliary training samples;

training, using the set of auxiliary training samples, a baseline interpretable model;

training, using the set of auxiliary training samples and the baseline interpretable model, an instance-wise weight estimator model;

for each auxiliary training sample in the set of auxiliary training samples, determining, using the trained instance-wise weight estimator model, a selection probability for the auxiliary training sample;

selecting, based on the selection probabilities, a subset of auxiliary training samples from the set of auxiliary training samples; and

training, using the subset of auxiliary training samples, a locally interpretable model.

12. The system of claim 11 , wherein training the instance-wise weight estimator model comprises:

determining a reinforcement signal; and

updating the instance-wise weight estimator model based on the reinforcement signal.

13. The system of claim 12 , wherein determining the reinforcement signal comprises:

determining first loss data for the locally interpretable model based on a loss function;

determining second loss data for the baseline interpretable model based on the loss function; and

determining the reinforcement signal based on the first loss data for the locally interpretable model and the second loss data for the baseline interpretable model.

14. The system of claim 11 , wherein selecting the subset of auxiliary training samples from the set of auxiliary training samples comprises, for each auxiliary training sample in the set of auxiliary training samples:

determining, based on the selection probability, a corresponding selection value indicating either selection or no selection; and

when the corresponding selection value indicates selection, adding the auxiliary training sample into the subset of auxiliary training samples.

15. The system of claim 14 , wherein determining, for each auxiliary training sample, the selection probability for the auxiliary training sample comprises:

receiving, at the instance-wise weight estimator model, a validation sample; and

determining, by the instance-wise weight estimator model, the selection probability based on the validation sample.

16. The system of claim 11 , wherein the operations further comprise, after training the instance-wise weight estimator model:

receiving, at the instance-wise weight estimator model, a validation sample;

generating, by the instance-wise weight estimator model, the selection probability for each auxiliary training sample in the set of auxiliary training samples; and

training the locally interpretable model based on the validation sample and the selection probabilities of each auxiliary training sample in the set of auxiliary training samples.

17. The system of claim 11 , wherein training the black-box model comprises minimizing a mean squared error loss function or a cross-entropy loss function.

18. The system of claim 11 , wherein each training sample in the set of training samples comprises training data and a given label.

19. The system of claim 18 , wherein generating the set of auxiliary training samples comprises, for each training sample in the set of training samples:

obtaining, from the trained black-box model, using the training data of the training sample, a predicted label for the training data;

combining the training data and the predicted label into an auxiliary training sample; and

adding the auxiliary training sample to the set of auxiliary training samples.

20. The system of claim 11 , wherein training the instance-wise weight estimator model comprises:

obtaining a set of probe training samples;

for each probe training sample in the set of probe training samples, determining a selection probability;

determining, based on the selection probability of each probe training sample, an instance-wise selection vector;

optimizing the locally interpretable model based on the instance-wise selection vector; and

updating the instance-wise weight estimator model based on loss data from the optimized locally interpretable model.