The key to generalization is abstraction. To understand DL, we need mathematical models of abstraction itself. An abstraction of abstraction
The surprising thing about deep learning isn't that our models can fit to basically any dataset, it is that they generalize --sometimes.
-
-
-
2deep4me
End of conversation
New conversation -
-
-
Is that only surprising about DL, or ML algos more generally?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Can we call DL a function approximator. A function whose output most of times is a distributed representation
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Imo more specifically, certain kinds of generalization are especially surprising; that DL can interpolate is less surprising.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
@Miles_Brundage I actually see the data fit bit as the biggest surprise, and generalization as a problem of incomplete data space -
because then you have a bigger bag of tricks, both generalization tricks and artificial data tricks
End of conversation
New conversation -
-
-
@Miles_Brundage but in a way yes, the most "powerful" ML in one sense is a hashtable, which fits the data perfectly every timeThanks. 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.