Disagree there. Nativism is useful for shortcuts, but comes with potential (hardcoded) dead-ends (stuck local optima, walled off solution spaces). So in my eyes, they are hacks that can yield interesting results short/mid term results, but also create hurdles on path to AGI.https://twitter.com/GaryMarcus/status/951048978988261376 …
-
-
Replying to @pwlot
Absolutely. Evolution is an extremely slow and unreliable (1.5Bn years stuck in local optima!) Monte Carlo style search for meta-learners, and results in a few dozen or hundred biases that speed up convergence for particular problem spaces. This can be better done on other ways.
1 reply 0 retweets 1 like -
Replying to @Plinz
Yep. Though, by Moravec, it's a hairy issue. The devil is in the details and all that. It hits all open questions and currently investigated avenues (consciousness, top down/bottom up learning, agency, data sets, modularity, embodiment, etc.)
1 reply 0 retweets 0 likes -
Replying to @pwlot
Most of which are red herrings. Even bottom-up/top-down might come down to optimizing recurrence. Embodiment has always been great for marketing and funding, but practically mostly bs.
1 reply 0 retweets 0 likes -
Replying to @Plinz
Maybe. " Even bottom-up/top-down might come down to optimizing recurrence.".. close to semantic sleight of hand, but sure...
1 reply 0 retweets 0 likes
What I mean is that bottom-up/top-down processing might possibly fall out naturally when you treat cognition is a regulation process that employs self-optimizing hierarchical function approximation.
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.