a year ago today I posted “deep learning: a critical appraisal”, asking whether DL might be approaching a wall, and be just one tool among many needed for AGI. hive mind: what progress has been made? what new limits have been identified? what other tools have been invented?
-
Progress towards (but not solved) - use of laws of physics as a prior - interal “world models” in RL - Nash equilib as learning (ie Tsetlin machine) - more general DNNs (ie Graph Networks) - better problem formulation (ie Neural ODE)
-
-
Best from.. DL world = [1] ‘physics prior’ *learned* & then applied with in RL scheme. CG world = [2] ‘physics prior’ encoded as heuristic for specific object & then used in RL scheme w/ goal of realistic behavior. [1] https://openreview.net/pdf?id=SklrrhRqFX … [2] https://arxiv.org/pdf/1804.02717.pdf …
-
Maybe this year we will see the methods of encoding of physics priors from the CG world (hand coded physics simulation) make its way into the RL/DL community working towards more general goals?
End of conversation
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.