Many people sharing this essay arguing that "computational scale beats clever new ideas". It takes for granted backprop, better activation functions, better learning methods, conv nets, better regularization techniques, etc etc. In other words, it seems to ignore the clever ideashttps://twitter.com/gdb/status/1106329741785653248 …
-
-
Of course, I don't disagree at all with the essay that computational scale buys you an incredible amount, and it's easy to under-estimate. But GANs & levering pawns (etc) weren't discovered with computational scale.
Show this thread -
Nice followup from
@gdb, making the point that it's a false dichotomy (scale versus cleverness):https://twitter.com/gdb/status/1106386231338762240 …Show this thread -
A better essay collecting some (correct!) examples in the same general direction is this paper by Banko & Brill: http://www.aclweb.org/anthology/P01-1005 … See eg this great graph, showing performance as a function of training data size. In this example: more data >> smarter algorithmpic.twitter.com/zwEmQSGMcP
Show this thread -
There is, IMO, a good paper to be written following this up, carefully understanding the relationship between scale and clever ideas.
Show this thread
End of conversation
New conversation -
-
-
good bit of research there...
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I take umbrage by characterizing Deep Blue as "just massive search." IMHO IBM cheated by reprogramming DB as the game was advancing https://www.technologyreview.com/s/400089/how-the-chess-was-won/ … See also the 44th move https://www.engadget.com/2014/10/23/fivethirtyeight-kasparov-deep-blue-ibm/ …pic.twitter.com/xhWldAtglz
Thanks. 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.