Interpretable AI: Building explainable machine learning systems by Ajay Thampi

Author:Ajay Thampi [Thampi, Ajay]
Language: eng
Format: epub
Publisher: Manning Publications Co.
Published: 0101-01-01T00:00:00+00:00

Figure 5.18 A saliency map using vanilla backpropagation

Let’s focus on the IDC-negative patch first. The original image of the patch is shown on the top left. The grayscale representation of the gradients obtained through vanilla backpropagation is shown on the top right. We can see pixels with various shades of gray in the image. Larger gradients have a higher intensity of gray or appear white. This is a great way to visualize what pixels the CNN is focusing on to predict that the image is IDC negative with a confidence of 82%. The salient, or important, pixels correspond to the dense lighter stains in the original image. Because gradients are shown at the pixel level, this is a much more fine-grained interpretation than LIME, where LIME focuses only on superpixels. Saliency maps are a great way to debug the CNN as a data scientist or engineer and also help the expert pathologist understand what parts of the image the CNN is paying attention to.

Let’s now look at the IDC-positive patch. The original image is shown on the bottom left, and the interpretation obtained through vanilla backpropagation is shown on the bottom right. We can see that a lot more pixels are lit up, which corresponds to the darker stains or the nuclei in the input image. This interpretation is a lot better than the LIME interpretation, where in LIME, the whole image was treated as a superpixel.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.