eduzhai > Applied Sciences > Engineering >

Explaining in Style Training a GAN to explain a classifier in StyleSpace

  • KanKan
  • (2) Download
  • 20210429
  • Save

... pages left unread,continue reading

Document pages: 10 pages

Abstract: Image classification models can depend on multiple different semanticattributes of the image. An explanation of the decision of the classifier needsto both discover and visualize these properties. Here we present StylEx, amethod for doing this, by training a generative model to specifically explainmultiple attributes that underlie classifier decisions. A natural source forsuch attributes is the StyleSpace of StyleGAN, which is known to generatesemantically meaningful dimensions in the image. However, because standard GANtraining is not dependent on the classifier, it may not represent theseattributes which are important for the classifier decision, and the dimensionsof StyleSpace may represent irrelevant attributes. To overcome this, we proposea training procedure for a StyleGAN, which incorporates the classifier model,in order to learn a classifier-specific StyleSpace. Explanatory attributes arethen selected from this space. These can be used to visualize the effect ofchanging multiple attributes per image, thus providing image-specificexplanations. We apply StylEx to multiple domains, including animals, leaves,faces and retinal images. For these, we show how an image can be modified indifferent ways to change its classifier output. Our results show that themethod finds attributes that align well with semantic ones, generate meaningfulimage-specific explanations, and are human-interpretable as measured inuser-studies.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×