eduzhai > Applied Sciences > Engineering >

Scientific Discovery by Generating Counterfactuals using Image Translation

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 12 pages

Abstract: Model explanation techniques play a critical role in understanding the sourceof a model s performance and making its decisions transparent. Here weinvestigate if explanation techniques can also be used as a mechanism forscientific discovery. We make three contributions: first, we propose aframework to convert predictions from explanation techniques to a mechanism ofdiscovery. Second, we show how generative models in combination with black-boxpredictors can be used to generate hypotheses (without human priors) that canbe critically examined. Third, with these techniques we study classificationmodels for retinal images predicting Diabetic Macular Edema (DME), where recentwork showed that a CNN trained on these images is likely learning novelfeatures in the image. We demonstrate that the proposed framework is able toexplain the underlying scientific mechanism, thus bridging the gap between themodel s performance and human understanding.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×