Are neural nets (the weights file) copyrightable?

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

Michel
Posts: 2272
Joined: Mon Sep 29, 2008 1:50 am

Re: Are neural nets (the weights file) copyrightable?

Post by Michel »

So it seems to me that the mechanism by which the GPL 3 works is as follows.

  • If you receive a copy of GPL 3 covered software you are not bound by its license. However if you want to _anything_ with it to which copyright law is applicable you have to accept the GPL 3 licence.
  • The GPL 3 puts restrictions on "covered works". The definition of covered work is extremely broad.
    A “covered work” means either the unmodified Program or a work based on the Program.
  • You may not like these restrictions or perhaps you are not in a position to meet them because of other licensing constraints. That's perfectly fine. This simply means you cannot accept the GPL 3 license and hence you cannot do anything with the software to which copyright law is applicable.
Ideas=science. Simplification=engineering.
Without ideas there is nothing to simplify.
Michel
Posts: 2272
Joined: Mon Sep 29, 2008 1:50 am

Re: Are neural nets (the weights file) copyrightable?

Post by Michel »

Note that even ChessBase admits that FF2 is based on Stockfish. So it is definitely a covered work in the sense of the GPL 3. Whether or not the GPL 3 applies to it depends on the aggregation exception. I assume the aggregation exception does not apply (since the situation is similar to dynamic linking) but I have not tried to parse it.
Ideas=science. Simplification=engineering.
Without ideas there is nothing to simplify.
Modern Times
Posts: 3550
Joined: Thu Jun 07, 2012 11:02 pm

Re: Are neural nets (the weights file) copyrightable?

Post by Modern Times »

Michel wrote: Sat Feb 27, 2021 9:40 am Note that even ChessBase admits that FF2 is based on Stockfish. So it is definitely a covered work in the sense of the GPL 3.
The Stockfish exe, yes. The NNUE, no. My opinion differs from yours. Who is right, only a court case will answer that.
Michel
Posts: 2272
Joined: Mon Sep 29, 2008 1:50 am

Re: Are neural nets (the weights file) copyrightable?

Post by Michel »

Modern Times wrote: Sat Feb 27, 2021 9:52 am
Michel wrote: Sat Feb 27, 2021 9:40 am Note that even ChessBase admits that FF2 is based on Stockfish. So it is definitely a covered work in the sense of the GPL 3.
The Stockfish exe, yes. The NNUE, no. My opinion differs from yours. Who is right, only a court case will answer that.
It is not a matter of opinion. It boils down to whether the aggregation exception applies to dynamic linking. The GPL FAQ says no. You can of course disagree with the GPL FAQ but then you are disagreeing with a professor at Colombia Law school....
Ideas=science. Simplification=engineering.
Without ideas there is nothing to simplify.
noobpwnftw
Posts: 560
Joined: Sun Nov 08, 2015 11:10 pm

Re: Are neural nets (the weights file) copyrightable?

Post by noobpwnftw »

It does not matter whether the NNUE is independently copyrightable. SF<->NNUE interaction is not "trivial" public interface, and there is no such "controlled interface" clause defined and explicitly allowing to link with an proprietary library in SF's license.

So in doing so(distributing a bundle of the two), those two need to be made license compatible, or you have that exception clause added, by the original author, and explicitly carried forward for any derivative. GPL FAQ is pretty clear on this matter.
syzygy
Posts: 5566
Joined: Tue Feb 28, 2012 11:56 pm

Re: Are neural nets (the weights file) copyrightable?

Post by syzygy »

Michel wrote: Sat Feb 27, 2021 6:46 am
syzygy wrote: Fri Feb 26, 2021 11:47 pm
syzygy wrote: Fri Feb 26, 2021 11:04 pmThe FSF's FAQ contains many untruths.
The following paper makes for interesting reading and comments on the absurdity of the FSF's position:
https://courses.cs.washington.edu/cours ... rksGPL.pdf

This article is about the GPL 2 and the wording of the GPL 3 is significantly different. In particular the GPL 3 no longer uses the terminology of derivative work.
GPLv3 is indeed worded differently, but this shows that whatever is in the FSF's GPL FAQ has to be taken with a big grain of salt.
I don’t see how a license cannot prohibit distribution in a larger package if certain conditions are not met. This is clearly what the GPL 3 attempts to achieve and its application to concrete cases (e.g. dynamic linking which is similar to the use of a NNUE net) is detailed in the FAQ. I agree that it is not always obvious how to extract the claims in the FAQ from the license itself but I rely on the fact that the GPL 3 was written by a Columbia law professor and has undergone substantial review before it was accepted.
But a reviewer such as Lawrence Rosen seems to agree with me that it is fine to package independent copyrightable works together even if they combine to make something more sophisticated.
syzygy
Posts: 5566
Joined: Tue Feb 28, 2012 11:56 pm

Re: Are neural nets (the weights file) copyrightable?

Post by syzygy »

Modern Times wrote: Sat Feb 27, 2021 7:12 am To me it is crystal clear that the NNUE distributed alongside the stockfish binary is in no way covered by the GPL. Sure I'm not a lawyer, and even the lawyers may disagree on it. No matter how many more pages this thread drags on the position won't be any clearer.
I think most people here agree on this now, even though they rely on GPL FAQ sections asserting the contrary :-)

The argument now is that, if you have a non-GPL regular SF NNUE net on your website, then you are violating the copyright on SF if you put up SF (with source) on your website as well.

Of course this whole discussion is academic, since there is no copyright on the NNUE net in the first place.
Michel
Posts: 2272
Joined: Mon Sep 29, 2008 1:50 am

Re: Are neural nets (the weights file) copyrightable?

Post by Michel »

syzygy wrote: Sat Feb 27, 2021 11:30 am
Michel wrote: Sat Feb 27, 2021 6:46 am
syzygy wrote: Fri Feb 26, 2021 11:47 pm
syzygy wrote: Fri Feb 26, 2021 11:04 pmThe FSF's FAQ contains many untruths.
The following paper makes for interesting reading and comments on the absurdity of the FSF's position:
https://courses.cs.washington.edu/cours ... rksGPL.pdf

This article is about the GPL 2 and the wording of the GPL 3 is significantly different. In particular the GPL 3 no longer uses the terminology of derivative work.
GPLv3 is indeed worded differently, but this shows that whatever is in the FSF's GPL FAQ has to be taken with a big grain of salt.
I assume that because of the ambiguities in the GPL 2 they decided to get serious legal council for the GPL 3. By using their own definition of "covered work" they circumvented having to rely on copyright law to define a derived work.
Ideas=science. Simplification=engineering.
Without ideas there is nothing to simplify.
gonzochess75
Posts: 208
Joined: Mon Dec 10, 2018 3:29 pm
Full name: Adam Treat

Re: Are neural nets (the weights file) copyrightable?

Post by gonzochess75 »

syzygy wrote: Sat Feb 27, 2021 11:30 am But a reviewer such as Lawrence Rosen seems to agree with me that it is fine to package independent copyrightable works together even if they combine to make something more sophisticated.
It does seem Lawrence Rosen agrees with you, but his comments are in the context of the GPLv2 and not the GPLv3 right?
syzygy
Posts: 5566
Joined: Tue Feb 28, 2012 11:56 pm

Re: Are neural nets (the weights file) copyrightable?

Post by syzygy »

Michel wrote: Sat Feb 27, 2021 11:47 am
syzygy wrote: Sat Feb 27, 2021 11:30 am GPLv3 is indeed worded differently, but this shows that whatever is in the FSF's GPL FAQ has to be taken with a big grain of salt.
I assume that because of the ambiguities in the GPL 2 they decided to get serious legal council for the GPL 3. By using their own definition of "covered work" they circumvented having to rely on copyright law to define a derived work.
A covered work is unmodified SF or "a work based on SF", which is the modified FF2 SF.

Section 4 is about distributing SF.
Section 5 is about distributing FF2 SF. It states that you are allowed to distribute FF2 SF if (condition c) you license the entire FF2 SF under the GPLv3.

The "aggregate" section of section 5 then, at worst, states that a zip file with FF2 SF and FF2 NNUE is not an aggregate, depending on interpretation. But no legal consequence is attached. It does not say that FF2 SF is no longer a covered work by itself if it is distributed as part of a compilation which is not an aggregate.