Direct feedback alignment for training neural networks: https://arxiv.org/abs/1609.01596 - can train NNs with random projections of the error signal.
-
As measured against absolutely atrocious backprop baselines in most cases.
That it works at all is interesting. How well it works is secondary, since this is not meant to be practical...
-
-
Is it that interesting? Geoff pointed out > a year ago that disjoint forward/backward paths work just fine.
-
and we all know that implementing your backwards pass with a bunch of bugs doesn't prevent learning. Still...
End of conversation
New conversation -
-
-
Given feedback alignment works, that this works isn't surprising. No mention of ResNets or "deeply supervised", both quite close.
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.