perfectly said. Whether we look at SQuAD or bABi or MCtest or any other test sets, a little deviation between test set and training set and the data based NLP will start going crazy... However soon data based machine learning will be replaced by textbook learning... #wait4vicki
This nails my central point — deep learning is best only if test data are drawn from same distribution as training data — and points to a provocative 2006 article (new to me, still relevant to the field) that shows this is a serious problem in the real world. thanks @MaxALittle !https://twitter.com/maxalittle/status/957976556273577984 …
-
-
-
Imagine an application where you simply describe what a dog looks like in plain simple english and the system is able to recognize a dog in picture... The required tech for that is coming soon...
#wait4vicki
End of conversation
New conversation -
-
-
Welcome! To be fair to deep nets, this applies to pretty much any classical "Kaggle competition"-like supervised learning setup, maximizing some single performance metric of mapping inputs to outputs. The danger lies in getting too obsessed in just lowering the error metric.
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.