Gradient descent will take any shortcut available to map inputs to targets. Human perception works differently: it starts from a different input (embodied stream vs static images), it doesn't have a target for each input, and it isn't trained with SGD.https://www.quantamagazine.org/where-we-see-shapes-ai-sees-textures-20190701/ …
-
This reminded me of a small experience I had when I was preparing a training on the MNIST dataset while having my 3-year-old son next to me: while the NN failed at guessing inverted images, 90 degrees rotated ones, my son guessed every single one of them.
By default we store visual representations in a symmetry-invariant way; letter orientation is something we have to learn, typically from ages 5 to 7. It's common for 5 y.o. to confuse d/b, write inverted letters, or even write in full mirror writing.
-
-
Indeed!! I later found out that he does not know the difference between 6 and 9. He learnt one of them and he stored it as you mentioned as symmetry-invariant and then when the other one came, he just considered it as the same as the first one.
-
It's really quite impressive -- for an adult it would take significant mental effort to draw the mirror version of a complex character, but children do it spontaneously & effortlessly after only looking at a few correctly-oriented examples. E.g. Japanese: https://matome.naver.jp/m/odai/2139156029049871901/2139156546952734803 …
End of conversation
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.