The second point is exemplified by "mirror writing" seen in kids learning to write - they store letter shapes in an orientation-indep way.pic.twitter.com/I3SCaq84TW
Show this thread
Human visual representations are not only translation-invariant (like CNNs), but also symmetry-invariant, & invariant for small rotations.
-
-
-
Wish I had more characters to talk about this...
Show this thread
End of conversation
New conversation -
-
-
They're not symmetry or rotation invariant -- doing the mirroring or rotation takes time.https://en.m.wikipedia.org/wiki/Mental_rotation …
-
Small rotations don't affect recognition / generation, but large rotations do. As for symmetry, the mirror effect is very well documented
- Show replies
New conversation -
-
-
So the reason I was going in that direction is, one way to get invariance is to use invariant operators, but another is to canonicalize.
-
Attention mechanisms for example can be used to center or align images/speech but you can also do it with muscle movement (tilt head, etc).
End of conversation
New conversation -
-
-
I bet, by covering oneself under the blanket of behaviourism like this, one can't create a real AI.Just some cheap tools
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
How can CNNs help with handwriting (not just digits) recognition
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Is there a way to do rotational convolution? Other than something like data-augmentation?
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.