I have mentioned the NTM only as a proof of concept (symbolic operations in NN). I don't think anyone but individual cells use a TM design for practical computations.
-
-
i talk about universal function approximation vs symbol manipulation in chapter two of algebraic mind; too subtle for twitter, perhaps
1 reply 0 retweets 1 like -
Replying to @GaryMarcus @Plinz
If you don't want to buy a book you can also check out http://www.psych.nyu.edu/gary/marcusArticles/Marcus%202013%20treelets%20%5Bproofs%5D.pdf … if you haven't. There are advantages, disadvantages, tradeoffs between explitcitly vs implicitly symbolic, discrete vs continuous and search/sampling vs optimization. Lots of people don't acknowledge
1 reply 0 retweets 2 likes -
And the idea that GOFAI ideas didn't lead anywhere is the biggest myth. Every database, verifier, WaveFunctionCollapse,symbolic algebra,type system, SMT is argument against it. In my case I could write yet another image analogy tool or I can write something that helps teach math
1 reply 0 retweets 1 like -
Replying to @sir_deenicus @GaryMarcus
I would not make any such claim. I also don't think that the ML folks deny that we learn how to reason symbolically. The strongest claims I hear are more along the lines that symbolic reasoning happens when the functions we approximate become low dimensional and discrete.
1 reply 0 retweets 1 like -
Replying to @Plinz @sir_deenicus
This simply rewrites history. See for example last sentence of Lecun, Hinton, Bengio in Nature 2015, or Rumelhart and McClelland’s declaration that symbol-manipulation is not essence of human cognition. Or Hinton’s “aeitherial symbols” talk.
1 reply 0 retweets 1 like -
Replying to @GaryMarcus @sir_deenicus
I don't see this as a contradiction. The essence of human cognition may better be described by the mechanisms that enable perception. I think it may be more plausible to extend perception into symbol manipulation than the other way around.
1 reply 0 retweets 2 likes -
Replying to @Plinz @sir_deenicus
? You said that the ML folks dont deny that we (learn) how to reason symbolically. They do (or did; Yann seemed to backtrack some today).
1 reply 0 retweets 0 likes -
Replying to @GaryMarcus @sir_deenicus
I thought the difference between your and their position is that they claim that symbolic reasoning is learned, while you seem to suggest that it is innate?
1 reply 0 retweets 2 likes -
Replying to @Plinz @GaryMarcus
Ah,that's a super reasonable position. I'm towards innate as it seems to have occured only in humans & not again before humans came to dominate. Seems an accounting trick if only human brains can learn this non-innate thing. Some (VWF) areas seem specialized but in general way
1 reply 0 retweets 0 likes
The obvious counter argument is that organismic evolution wrt minds constitutes a worst case unprincipled blind algorithm for hyperparameter optimization, which we should be able to outperform in all sorts of ways.
-
-
Replying to @Plinz @GaryMarcus
But this counter is in the wrong direction. It's not about what we can implement but that it seems brains that can recursively handle symbols require particular existing structures. It is not a natural aspect of all brains despite their being (in principle) Turing equivalent
0 replies 0 retweets 0 likesThanks. 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.