Generating saliency masks for inputs of models using saliency metric

1. A system, comprising a processor to:

receive an input and a model trained to classify inputs;

iteratively generate a perturbed input that optimizes a saliency metric including a classification term that measures an amount of change in a classification of the model with regards to a class given a change in the input, a sparsity term, and a smoothness term, while keeping parameters of the model constant;

detect that a predefined number of iterations is exceeded or a convergence of values of the perturbed input; and

generate a saliency mask for the input based on a perturbation of the perturbed input in response to detecting the predefined number of iterations is exceeded or the convergence.

2. The system of claim 1 , where the perturbation from the input is transformed and thresholded to generate a binary saliency mask.

3. The system of claim 1 , wherein the classification term represents destructiveness of a saliency region with respect to an object class.

4. The system of claim 1 , wherein the sparsity term is to reduce the total number of values changed in the perturbed input relative to the input.

5. The system of claim 1 , wherein the smoothness term is to increase clustering of components in the saliency mask.

6. The system of claim 1 , wherein the perturbed input is generated using a loss function that is an approximated version of the saliency metric.

7. The system of claim 1 , wherein the saliency metric comprises an adversarial perturbative explanation metric based on smallest sufficient region or smallest destroying region.

8. A computer-implemented method, comprising:

receiving, via a processor, an input and a model trained to classify inputs;

iteratively generating, via the processor, a perturbed input by transforming a perturbation of the input to optimize a saliency metric including a classification term that measures an amount of change in a classification of the model with regards to a class given a change in the input, a sparsity term, and a smoothness term;

detecting, via the processor, that a predefined number of iterations is exceeded or a convergence of values of the perturbed input; and

generating, via the processor, a saliency mask for the input by thresholding the transformed perturbation.

9. The computer-implemented method of claim 8 , wherein iteratively generating the perturbed input comprises clipping values of the perturbed input to constrain the perturbed input within a range of original applicable values from which the input is sampled.

10. The computer-implemented method of claim 8 , comprising generating an improved perturbed input with non-zero values only inside the saliency mask, the improved perturbed input to optimize the classification term, and generating the saliency mask by thresholding an updated transformed perturbation of the improved perturbed input.

11. The computer-implemented method of claim 8 , wherein iteratively generating the perturbed input comprises iteratively modifying the perturbed input based on a gradient of a loss with respect to a previous perturbed input of the adversarial example generator while holding parameters of the model constant.

12. The computer-implemented method of claim 8 , comprising classifying the input via the model and providing the saliency mask with a classification of the input.

13. The computer-implemented method of claim 8 , comprising locating an object associated with a classification of the input in the input based on the saliency mask.

14. The computer-implemented method of claim 8 , wherein iteratively generating the perturbed input comprises back propagating a loss through the trained model.

15. A computer program product for generating saliency masks, the computer program product comprising a computer-readable storage medium having program code embodied therewith, wherein the computer readable storage medium is not a transitory signal per se, the program code executable by a processor to cause the processor to:

receive an input and a model trained to classify inputs;

iteratively generate a perturbed input that optimizes a saliency metric including a classification term that measures an amount of change in a classification of the model with regards to a class given a change in the input, a sparsity term, and a smoothness term, while keeping parameters of the model constant;

detect that a predefined number of iterations is exceeded or a convergence of values of the perturbed input; and

generate a saliency mask for the input based on a perturbation of the perturbed input in response to detecting the predefined number of iterations is exceeded or the convergence.

16. The computer program product of claim 15 , further comprising program code executable by the processor to clip values of the perturbed input to constrain the perturbed input within a range of original applicable values from which the input is sampled.

17. The computer program product of claim 15 , further comprising program code executable by the processor to further iteratively generate an improved perturbed input with non-zero values only inside the saliency mask, the improved perturbed input to reduce the classification term, and generate the saliency mask by thresholding values of the improved perturbed input at a predetermined threshold.

18. The computer program product of claim 15 , further comprising program code executable by the processor to iteratively modify the perturbed input based on a gradient of a loss with respect to a previous perturbed input while holding parameters of the model constant.

19. The computer program product of claim 15 , further comprising program code executable by the processor to classify the input via the model and provide the saliency mask with a classification of the input.

20. The computer program product of claim 15 , further comprising program code executable by the processor to locate an object associated with a classification of the input in the input based on the saliency mask.