I would bet against current backprop-based techniques in the long run.https://twitter.com/michael_nielsen/status/669978274701942785 …
-
.
@fchollet What do you think is a good direction for replacements?
@michael_nielsen Information geometry / topology will give us the theoretical foundation for building "optimal" learning algorithms.
-
-
@fchollet@michael_nielsen any recommendations there? I'm reading Amari and Nagaoka (slowly) -
@wxswxs@michael_nielsen That's a great start. Otherwise there have been a number of interesting ML papers on Riemannian manifolds.
End of conversation
New conversation -
-
-
@fchollet@michael_nielsen I can't see local (even global sometimes) optimization without doing gradient descent in *some * space. -
@edersantana@michael_nielsen learning problems can be formulated as numerical optimization problems, but that is not the only way. - Show replies
New conversation -
-
-
@fchollet btw. agree, info geo seems like the way to go. Why i will start to call "Design" from here on out "Information Geography" ;-)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.