@fchollet I've always been suspicious of it; I think it was Zeiler who showed strided convolutions were better/faster at reducing dimensions
Convnets: at this point I am 90% sure we should get rid of max pooling / average pooling altogether.
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
@fchollet What would you replace them with? - Show replies
New conversation -
-
-
@fchollet I'm apparently behind the times on this. Is there a reference or explanation where I can learn why? Just longer strides instead? -
@dan_s_becker in deep learning, for now, there are no explanations : ) There have been a few papers about it though - Show replies
New conversation -
-
-
@fchollet Do you also propose something other then FCs at the end? w/o Pooling those are huge. That would also change a key CNN assumption. -
@colinlea striding your convolutions accomplishes the dimensionality reduction http://arxiv.org/abs/1412.6806 - Show replies
New conversation -
-
-
@fchollet +1. E.g. http://arxiv.org/abs/1412.6806 . In particular for segmentation they are a bad thing to use...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.