Asking to people who believe Leela NN is a book, what they think about SF NN now?

[Ovyron]

You've gone from arguing the idea, to arguing the semantics.

"Leela NN is a book" has always been about semantics. It depends of your definition of "book."

Your side just can't refute the argument that if Leela could memorize moves, she could play them from memory instantly.

Checkmate.

[jhellis3]

Somehow a leela style net that never saw positions with more than 18 pieces was able to “memorize” opening moves. Yet somehow people arguing this tired old argument are unable to memorize this refutation.

I'm quite sure myself and others have pointed out the logical flaw in this reasoning, yet you continue to post it.

B containing b and A containing b does not imply that A is a subset of B.

[AndrewGrant]

You've gone from arguing the idea, to arguing the semantics.

"Leela NN is a book" has always been about semantics. It depends of your definition of "book."

Your side just can't refute the argument that if Leela could memorize moves, she could play them from memory instantly.

Checkmate.

No, lol. You boldly claimed that NNs don't memorize things. I refuted that by providing a paper that demonstrates under broad conditions that NNs can achieve the same error rate with correctly labeled data as they can with randomly labeled data, which is akin to memorization. I'm not making any statements about Leela, I'm pointing out that your presuppositions of what an NN can and cannot do is wrong.

Is Leela memorizing openings? I don't know. Depends on the data shes given.
Is it possible for Leela to have memorized the openings via training? Yes

Words are important. I choose mine carefully.

[dkappe]

Somehow a leela style net that never saw positions with more than 18 pieces was able to “memorize” opening moves. Yet somehow people arguing this tired old argument are unable to memorize this refutation.

I'm quite sure myself and others have pointed out the logical flaw in this reasoning, yet you continue to post it.

B containing b and A containing b does not imply that A is a subset of B.

Ah, if only this argument was an exercise in elementary set theory. It’s not. Yet you still don’t get it.

[jhellis3]

I get it quite well. And I have explained it explicitly before and am fine to do it again.

Just because one can train a NN to play "theory" without ever seeing an opening position does not mean that a NN when provided such data and the features/tools to identify and make use of such data will not do so. Period.

You think that what you have is a counter example. It isn't.

And quite frankly, you are not doing your reputation any favors here... not that it is any of my business.

[dkappe]

I get it quite well. And I have explained it explicitly before and am fine to do it again.

Just because one can train a NN to play "theory" without ever seeing an opening position does not mean that a NN when provided such data and the features/tools to identify and make use of such data will not do so. Period.

You think that what you have is a counter example. It isn't.

And quite frankly, you are not doing your reputation any favors here... not that it is any of my business.

You are clearly not familiar with how convolutional neural nets function. You remind me of that guy who prattles on and on about Hopfield networks, believing that just because they have “neural network” in their name, they have anything to do with ResNets.

[jhellis3]

Ah, the ad hominem, the bastion of the well reasoned....

[Ovyron]

Words are important. I choose mine carefully.

No, you don't. You're using the word "can", but should be use the word "could":

NNs could solve world hunger, and they could heal any illness, and they could help you build a time machine.

And NNs could memorize opening positions, and their best moves. NNs could be a book. Sure.

But I was very careful with my word choice, all the times I mentioned NNs, I said "that NN" (you can go back an see how often I used the word "that") not "an NN", I was specifically talking about *that* NN.

And THAT NN does NOT memorize openings.

You're talking about what it's capable of doing, sure, anyone can program an NN to do anything, memorizing opening positions could be trivial (you just let it overfit the data), I'm taking about what it DOES, in the PRESENT, at the CURRENT TIME.

And it's not memorizing opening positions, nor doing anything akin to it. It could do it, and then it'd be a book, but it's not.

[Alayan]

It's pointless to argue with people who don't care about making gross logical mistakes in their reasoning.

[dkappe]

Ah, the ad hominem, the bastion of the well reasoned....

Pot, kettle and so on, Mr. Reputation.

Unfortunately you’ve got a vacuous hypothesis on your hands. If it isn’t vacuous, it must be possible, with the proper evidence, to disprove it. So, Mr. Reputation, what kind of evidence would disprove your hypothesis?