Gary—do you think "operations [over variables] a la logic" and "symbolic techniques" can be learned within the differentiable framework, or do you imagine that they are hard-wired or are very strong structural priors?
Retweeting my own tweet, because I don’t want it get lost in the stack of discussions about less relevant issues. THIS is what the debate w @ylecun is really about: whether or not we need machinery for performing operations over variables, in addition to vectors and gradients.https://twitter.com/GaryMarcus/status/1071452576074280960 …
-
-
-
Symbolic operations can certainly be learned by differentiable frameworks. But by default they learn propositional rules and not first-order rules (with universally quantified variables). Neural networks suffer, by default from what McCarthy called the "propositional fixation".
- 1 more reply
New conversation -
-
-
The vectors are discrete components of the symbols being manipulated, but we don't need to model consciousness at the hardware level.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Gary -- Yann wants to numerify everything. By doing that there is a significant semantic information loss.
- 1 more reply
New conversation -
-
-
This Tweet is unavailable.
- 1 more reply
-
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.