Framed properly, it does detect what you posted: https://imgur.com/a/0wdqPea. Positively identifying abstract geometric loss while rejecting non-loss text and drawings was something I had to explicitly optimize the training sets for.
NEAT. I rate this (1/2) / (2/50).
In the screenshots shown in the link you can see that it detects plenty of "disguised" Loss images.
By the way, cool app eigenloss, way more accurate than I'd have expected.
Thanks lccarrasco! It does work surprisingly well for such a tiny CNN. Props to the MobileNet team at Google.