Since we have image model support, it would be nice to have some visualizations of what areas of the image are important.
I don't know about the front end part, but I was thinking once training has finished there would be an additional section un Deep Learning Results. There, the user can select the image they are interested in, and see visualizations of significant regions.
GradCAM or Guided BackProp can be used for that. Like this:
Since we have image model support, it would be nice to have some visualizations of what areas of the image are important.
I don't know about the front end part, but I was thinking once training has finished there would be an additional section un Deep Learning Results. There, the user can select the image they are interested in, and see visualizations of significant regions.
GradCAM or Guided BackProp can be used for that. Like this:
This can either be implemented manually (see https://arxiv.org/abs/1610.02391) or using a library called
captum
.