Twitter | Search | |
Daniel Roy Jul 17
Ambivalent about the name of this award.
Reply Retweet Like
Blake Richards Jul 17
Replying to @roydanroy
~Isn't it ironic...~
Reply Retweet Like
Venkat Ramaswamy Jul 17
Why?
Reply Retweet Like
Blake Richards Jul 18
Minsky set the field of NNs back by years with his Perceptrons book.
Reply Retweet Like
Venkat Ramaswamy Jul 18
Was it his fault? He (& Papert) only pointed out the truth, with respect to limitations of single perceptrons, didn't he? Without the book, wouldn't it just have been another decade of delusion, followed by the inevitable winter? I frankly don't get the Minsky hate at all.
Reply Retweet Like
Blake Richards
I disagree. At the time, there was already work on MLPs, and M & P undersold their value. Based on some things I’ve heard about their feelings re Rosenblatt, it sounds like their book was not aiming for balance, it was aiming for a knock down.
Reply Retweet Like More
Venkat Ramaswamy Jul 18
I don't know. Perhaps the hate is directed at him, with the benefit of hindsight. If symbolic systems had worked out, for example, most of us would be calling him a genius today for the book.
Reply Retweet Like
Grace Lindsay Jul 18
oooo, gotta any sources on this gossip?
Reply Retweet Like
Daniel Roy Jul 18
But they didn’t. And when they do, they’ll bear little resemblance.
Reply Retweet Like
Blake Richards Jul 18
Ha! I've been told by one academic in the field, who I'll leave nameless (sorry, don't wanna out my source), that M & P found Rosenblatt's trumpeting of perceptrons irritating.
Reply Retweet Like
Blake Richards Jul 18
Yes, it's true that if symbolic systems had worked the post-hoc perception would be quite different. But at the same time, I think it has to be admitted that they set neural networks research back due to the way they framed things, and they didn't need to do that.
Reply Retweet Like
Venkat Ramaswamy Jul 18
Interesting debate. Even if folks had done the chain rule and gotten backprop working in the 70s, it didn't hit real paydirt until a few years ago due to processing power and datasets catching up, so I'm still on the fence about them materially setting things back.
Reply Retweet Like
Arnab Bhattacharyya Jul 18
Well, for complexity theorists, that book is heroic :) M & P proved the first unconditional lower bound on the degree of polynomial threshold functions! (Degree required to represent AND of two majorities is super-constant.)
Reply Retweet Like
Daniel Roy Jul 18
Maybe machine learning is good evidence that worst case analyses aren't relevant.
Reply Retweet Like
Arnab Bhattacharyya Jul 18
Agreed.
Reply Retweet Like