Neural networks basically just learn to classify textures.https://twitter.com/karpathy/status/1091813185995358208 …
-
To me that would be at odds with how effortless it is to eg watch cartoons, interpret art, etc. that said, I’ve seen a few examples where ConvNets also possess this ability but to a lower extent, I suspect because they are allowed to be lazy in their training data.
-
-
A compelling paper on this topic (https://arxiv.org/abs/1811.12231 ) demonstrates this is a case of a bias towards texture, not a limitation on shape. Compare to SGD for polynomial regression, which learns low freq first. If u use a diff basis (laguerre eg) you don't.
-
If you exclude texture as an option, say w/ style transfer but same labels, then net learns shape *and* is more robust.
- 2 more replies
New conversation -
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.