This is not what he said here. Every stepwise constructive or evolutionary process is in a sense an optimization via a gradient based method, so DL is a sufficiently broad methodology, but you'll still need to set up the system performing the optimization or meta optimization.
-
-
Replying to @Plinz @IntuitMachine
So, DL is an architecture AND a methodology AND the mechanism behind all of evolution? Next, you’ll tell me that is also serves to unify the theory of gravity with the standard model of physics!
@GaryMarcus1 reply 0 retweets 4 likes -
No, my understanding is that Yann LeCun thinks of DL as automating the optimization of a program for computing a complex function. Depending on the optimization method, the program or its generator function will often be differentiable.
1 reply 0 retweets 2 likes -
Indeed; I was simply building on what you said: “Every stepwise constructive or evolutionary process is in a sense an optimization via a gradient based method.”
1 reply 0 retweets 1 like -
It does not follow that each of these processes is also DL, but that DL (understood as a methodology) is not obviously unable to reproduce them
1 reply 0 retweets 1 like -
Indeed. Which is why I asked the question earlier, what is DL not?
1 reply 0 retweets 1 like -
Handcrafted solutions
2 replies 0 retweets 1 like -
does that mean no priors? isn't convolution a handcrafted prior? wasn't the structure of AlexNet handcrafted?
1 reply 0 retweets 2 likes -
Yes, but the difference is that DL is moving the handcrafting up one layer: classical AI mostly relied on writing algorithms, current AI mostly relies on writing algorithms that discover algorithms. Perhaps the next wave will be fueled by meta-learning, i.e. 3rd order programming
3 replies 0 retweets 2 likes -
i asked three questions; i don't quite understand how this answers any of them or which it is intended as an answer to. there's LOTS of handcrafting in deep learning that you seem to be ignoring. Some of the handcrafting may have moved up a later but there is plenty left.
3 replies 2 retweets 7 likes
Treaded discussion is hard on twitter. I agree that many DL solutions are hybrids (and don't think that this is a contested position). Automating architecture search/modification is only just beginning. But NLP and computer vision have basically dumped decades of prior work.
-
-
It's the 'dumping of decades of prior work' that has upset many. But this is the nature of science. There's no guarantee, no matter how philosophically sound ones arguments, that one can be totally wrong. GOFAI is useful, but not useful for a General Intelligence solution.
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.