This is from the Amazon description of @GaryMarcus new book, though I'm guessing there's a high probability these are not precisely his words. Nonetheless, the same argument pops up in other places too.
-
Show this thread
-
Replying to @santoroAI
yes. of course the point of writing the book is that argument is too subtle to capture in a jacket description or a tweet.
1 reply 0 retweets 2 likes -
Replying to @GaryMarcus @santoroAI
It is so subtle I've never quite grasped it. 'Endowing machines with common sense and deep understanding' in the obvious way was tried for 50 years (it's the obvious thing to try first) and didn't work. What's different about your proposal now? (ps. I'll read the book...)
4 replies 0 retweets 0 likes -
Replying to @FelixHill84 @santoroAI
Yes, please read the book. But also compare the argument you just made - “it’s already been tried, and failed” — to what might have been applied to deep learning in 2010.
3 replies 0 retweets 3 likes -
Replying to @GaryMarcus @FelixHill84
From what I've read in recent critiques of DL, the arguments for "more than statistical analyses/pattern recognition" also tend to simply point at what DL can't do (yet). Is there something in your book that *doesn't* point out DL failures, and is different from F+P 1988?
2 replies 0 retweets 0 likes -
Replying to @santoroAI @FelixHill84
inter alia, we examine in unprecedented depth what the problems for DL are, focusing on two detailed case studies, AND we also subject *symbolic approaches* to considerable scrutiny, & then make specific proposals about what problems in common sense have highest priority.
1 reply 0 retweets 2 likes -
Ah - "unprecedented depth". Always works :-)
1 reply 0 retweets 2 likes -
it matters—even in my DL critical appraisal, nothing conveyed the richness of the mismatch between what (pure) deep learning is doing & what actually needs to be done. greatest emphasis in book is on articulating true nature of the problem.
1 reply 0 retweets 0 likes -
Yes, sure. It was just a joke - because "unprecedented depth" is often the reason the latest DL architecture beats SoTA
1 reply 0 retweets 1 like -
ha ha!
1 reply 0 retweets 0 likes
Gary+others, have you considered making a big table that would have rows as research papers or groups thereof and columns as attributes like “pure deep learning”, “hybrid”, “has symbol manipulation”, and so on? Curious what the columns would be for NTM/DNC, say, in your views.
-
-
Replying to @AdamMarblestone @GaryMarcus and
more columns: Unsupervised, Learning from single input, Continuous learning, Open-ended processing, Detects non-solvable scenarios
0 replies 0 retweets 2 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.