Tweeted By @wattenberg
Did your neural net say "zebra" because of pixel (17, 153)? Or because of the stripes?
— Martin Wattenberg (@wattenberg) July 13, 2018
TCAV interprets neural nets with high-level concepts, not low-level features. See https://t.co/OiyDaXznkS and https://t.co/rjzoGrTkAv cc @_beenkim @jmgilmer @Carryveggies @bengiswex @viegasf pic.twitter.com/K5DXqeos5h