Detailed perturbative heatmap explanations
Saliency heatmap generation is disclosed. A mask that corresponds to an image is iteratively passed through a model. At each iteration, the output mask is optimized via gradient descent to gradually turn off one or more pixels. The output mask is iteratively passed in order to minimize contributions of pixels and to maximize a class prediction. After generating a final output mask, a saliency heatmap is generated.
1 . A method comprising:
receiving an image at a classifier engine;
generating an initial mask of the image;
passing the initial mask through a model to generate an output mask;
iteratively passing the output mask through the model until the output mask satisfies a loss threshold, wherein the output mask becomes a final output mask;
determining a loss value at each iteration, wherein the loss value is a combination of a class loss and a pixel loss, wherein the class loss is a difference between 1 and a class prediction and wherein the pixel loss is a sum of values in the mask for a given optimization divided by a sum of values in the initial mask; and
generating a heatmap from the final output mask, wherein the heatmap is a saliency heatmap that provides an explanation for a classification of the image generated by the classifier engine.
2 . The method of claim 1 , further comprising gradually changing values of pixels in the mask that contributing less to the classification performed by the classifier engine to black, wherein the values of the pixels that are contributing less are reduced at each iteration.
3 . The method of claim 2 , wherein gradually changing the values of pixels that are contributing less to the classification to black includes gradually changing the values of the pixels to 0 using gradient descent, wherein only pixels in the final output mask that contribute to the classification have a value greater than 0.
4 . The method of claim 1 , further comprising performing gradient descent in each iteration.
5 . The method of claim 1 , wherein iteratively passing the output mask through the model is configured to minimize contributions of pixels while maximizing a class prediction.
6 . The method of claim 1 , wherein the heatmap is determined by multiplying the image with the final output mask.
7 . The method of claim 1 , further comprising performing a Relu function to generate the heatmap.
8 . A non-transitory storage medium having stored therein instructions that are executable by one or more hardware processors to perform operations comprising:
receiving an image at a classifier engine;
generating an initial mask of the image;
passing the initial mask through a model to generate an output mask;
iteratively passing the output mask through the model until the output mask satisfies a loss threshold, wherein the output mask becomes a final output mask;
determining a loss value at each iteration, wherein the loss value is a combination of a class loss and a pixel loss, wherein the class loss is a difference between 1 and a class prediction and wherein the pixel loss is a sum of values in the mask for a given optimization divided by a sum of values in the initial mask; and
generating a heatmap from the final output mask, wherein the heatmap is a saliency heatmap that provides an explanation for a classification of the image generated by the classifier engine.
9 . The non-transitory storage medium of claim 8 , further comprising gradually changing values of pixels in the mask that contributing less to a classification performed by the classifier engine to black, wherein the values of the pixels that are contributing less are reduced at each iteration.
10 . The non-transitory storage medium of claim 9 , wherein gradually changing the values of pixels that are contributing less to the classification to black includes gradually changing the values of the pixels to 0 using gradient descent, wherein only pixels in the final output mask that contribute to the classification have a value greater than 0.
11 . The non-transitory storage medium of claim 8 , further comprising performing gradient descent in each iteration.
12 . The non-transitory storage medium of claim 8 , wherein iteratively passing the output mask through the model is configured to minimize contributions of pixels while maximizing a class prediction.
13 . The non-transitory storage medium of claim 8 , wherein the heatmap is determined by multiplying the image with the final output mask.
14 . The non-transitory storage medium of claim 8 , further comprising performing a ReLu function to generate the heatmap.