adversarial images for non-computer vision https://twitter.com/41Strange/status/1217861694094864384 …
-
This Tweet is unavailable.
-
Replying to @johnregehr
A hill I am going to die on is that adversarial images and camouflage in nature are two entirely different things. Adversarial images achieve a change in classification label by small changes in input image. Painting the whole thing in a certain way is not a small change.
2 replies 0 retweets 15 likes -
Replying to @oe1cxw @johnregehr
are there examples from nature, though, of small changes which defeat classification? but they'd probably be too subtle for humans to notice them…
1 reply 0 retweets 0 likes -
Replying to @hikari_no_yume @oe1cxw
it's an interesting thought! but of course natural neural networks tend to be pretty robust in common situations, out of necessity
2 replies 0 retweets 2 likes -
Replying to @johnregehr @oe1cxw
yeah, there'd be a strong evolutionary pressure for robustness on both sides, so I guess even if subtlety works once, your descendents will be less so
1 reply 0 retweets 2 likes
I guess cases where things have just two dots somewhere on their back and it thus looks like eyes on a huge head would count. I think I have seen (pictures of) insects with this kind of camouflage before.
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.