It's the results you get that make your algorithm interesting, not how elegant your theory is or how much time you spent developing it
Show this thread
Hence, please, don't evaluate any new ML algorithm on MNIST (solely), ever
-
-
Instead, spend as much time as possible demonstrating that you are getting good results across a range of tasks/datasets as wide as possible
Show this thread -
Since getting these results is the whole point of what you are doing -- not an afterthought
Show this thread
End of conversation
New conversation -
-
-
We always say that human accuracy is an important measure to test models against. In MNIST It is easy for humans to get 100% accuracy.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
MNIST's best error in ConvNets e.g. is 0.23 from Ciresan (2012). This has made me wonder why we consider MNIST a solved problem then?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Agreed but Building and discussing new algorithm on a well understood dataset even though a toy set is great, remember ...
-
Entire CNN approach was built on MNIST
- Show 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.