Lots of issues with MNIST, but most of all, it is really not representative of CV tasks. Please at least try CIFAR10, of comparable size.https://twitter.com/goodfellow_ian/status/852591106655043584 …
-
Many good ideas will not work well on MNIST (e.g. batch norm). Inversely many bad ideas may work on MNIST and no transfer to real CV.
Besides, MNIST cannot serve as a benchmark to compare different ideas. But I suspect that's why people love it so much. Competing is hard.
-
-
On real datasets, you often need state-of-the-art results to prove your approach (which is unfortunate IMO). On MNIST, you don't have to.
-
That's the real problem imo , SITS results grab attention while non-SOTA papers are generally ignored.
End of conversation
New conversation -
-
-
Do people still take these papers (which show their results only on MNIST) seriously?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Is notMNIST (http://yaroslavvb.blogspot.com/2011/09/notmnist-dataset.html …) better in that respect? (BTW: plans to include it in keras.datasets?)
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.