The twist of the Clever Hans story is that horses do actually have a concept of numbers (like most mammals). Just because a learning system takes a shortcut to perform a task w/o leveraging the relevant generalizable ability doesn't mean it doesn't possess it or can't acquire it
-
-
I think something similar too--after all, a "deep neural network" these days is just a matrix based computer program format with certain restrictions on control flow over numeric representations--but, I put the blame on loss functions. SGD is a search strategy, true but
-
I think the limitations come more from how guidance is provided on how to navigate that space via loss & problem formulation. So if you replaced SGD, you'd more be changing navigation & replacement need not be strictly better.
- Show replies
New conversation -
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.