Why spoze innateness is in the form of priors? That assumes learning by a type of enumerative induction/Bayes updating rather than conjecture/refutation. I wonder if AI would look different in the hands of a falsificationist/corroborationist. Stat inference looks very different.
Not so. I think the right priors could be evolved in silico, but it’s a hard task; it took evolution roughly a billion years to do it right, and it evolved programs, rather than connection weights. Discussed at length in my book The Birth of the Mind.https://twitter.com/plinz/status/968521519059136519 …
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I feel your argument doesn't address the point of
@Plinz. Evolution as an abstract optimization process can be sped up in simulation. And a neural network model is Turing complete, can represent any program as long as sufficient space is given for storage. Right? -
I agree with you that evolutionary processes in nature are discovering models that are far more sophisticated than current abstractions in software. But I believe this is because we don't know very well the nature of evolutionary processes. May be there are many things unknown.
End of conversation
New conversation -
-
-
How broad and how deep of a training space would you need to establish meaningful prior weights around spatial navigation, anatomical dynamics, facial features, emotional signaling, language, tone, family relationships...?
-
...and what would that loss function look like?
End of conversation
New conversation -
-
-
You can frame this in terms of pros and cons of natural evolution versus "artificial evolution". While the latter has cons, I'd also look closely to the cons of the former.
@PlinzThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Autopoiesis at the mercy of environment and its selection pressures. Being stuck in local optima is very different for the two, in a way that has everything to do with the time frame you mention.
@PlinzThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I would see a big difference between ad-hoc hand picked cost functions and the vastly richer and more information dense physics of the natural environment. One embodies much more info than the other.
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.