“Universal adversarial perturbations” seems most dramatic ML result in years; if so, not getting deserved attention https://arxiv.org/pdf/1610.08401v1.pdf …
-
-
Replying to @Meaningness
Result: there is a *fixed*, human-invisible map you can add to *any* image, and it renders it unclassifiable by multiple DL systemspic.twitter.com/s4zMsPy6Jy
22 replies 490 retweets 650 likes -
Replying to @Meaningness
I have long suspected that DL image classifiers depend mainly on texture and maybe color, making much less use of shape than vertebrates do…
6 replies 44 retweets 98 likes -
Replying to @Meaningness
Convolutions are obviously going to be good at picking up textures (if you have enough of them). Hard problem in vision always was…
1 reply 0 retweets 0 likes -
Replying to @Meaningness
3D shapes, when projected in 2D image, are entirely different depending on rotation. Worse when object is flexible or articulated.
1 reply 0 retweets 2 likes -
Replying to @Meaningness
I would guess DL systems can address shape problem only by brute force, i.e recognizing each pose separately, from examples of that pose.
1 reply 0 retweets 1 like -
Replying to @Meaningness
Coupling a convolutional texture-classifier with a 3D shape-from system, to handle rotation and pose, might win big.
5 replies 1 retweet 2 likes -
Replying to @Meaningness
You might be able to cheat just by being good enough at image segmentation?
2 replies 0 retweets 1 like
That’s certainly an important part at least! I was working on segmentation in early 90s but computers were too #%^*+ slow
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.