Interesting recurring debate! Some of your critique doesn't seem to hold up against modern models: the idea that input-output representations are doing the heavy lifting. Pixel-to-torque systems are routinely trained. Surely that counts as learning an internal representation?
the history of neural networks and the innateness question is now officially repeating itself. what i said once upon a time about the just cited (& worth reading) Rethinking Innateness;:. http://www.psych.nyu.edu/gary/marcusArticles/marcus%201998%20cognition.pdf …https://twitter.com/drjamesbooth/status/1002283208657068033 …
-
-
-
My emphasis, then and now, was on what I called in 1998 “genuine abstraction” and “featurewise indepedence”; it was clear you could generalize within a training space, but also that (certain kinds of) networks would have trouble extrapolating. Don’t think that’s changed.
- 1 more reply
New conversation -
-
-
Is there any neuroscience research to point to the existence of strongly interconnected synapses that are created during neonatal brain development? I'm assuming this would be incredibly difficult to study.
-
Not individual synapses, but the entire circuit is quite stereotypical in insects. In mammals, less so, but you still have stereotypical functionality -- same reflexes, etc, and the overall brain architecture is set up even in the absence of neural firing.https://twitter.com/generuso/status/942496009262944257 …
- 2 more 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.