Are neural nets (the weights file) copyrightable?

[towforce]

Apologies if someone else has already said this (I haven't read the whole thread), but it seems obvious to me that a set of NN weights can be copyrighted because:

1. I can copyright work no matter how obvious it is.

If nobody else has written "Grass is green, sky is blue and tree bark is brown" then I can copyright that expression (though such a copyright would not obviously be very valuable).

2. I can copyright work no matter how stupid it is.

If I can prove that I originated this work, then I could copyright "When iron is pure enough, then in its solid form at room temperature it will float in air because it's lighter and less dense than air at sea level."

One caveat: if your work is an "invention", you might need to patent it rather than copyright it, but I don't think you'll get a patent for a NN that's bigger than NNUE's but smaller than LC0's (though patents at that level of stupidity have been awarded).

A formula cannot be copyrighted - link - but I don't think that a set of NNUE weights would fall into that category.

I think that a set of NN weights would be like "chess advice": if I were to write, "Take your opponent's pieces, promote your pawns, and then trap your opponent's king", if nobody else had said that using those words, then I would be able to copyright them. Precedent: a lot of the ideas for Dan Brown's book "Da Vinci Code" came from "Holy Blood, Holy Grail", but the authors of the original book failed in their attempt to sue because no actual expressions from the original had been used - just ideas.

[Michel]

A formula cannot be copyrighted - link - but I don't think that a set of NNUE weights would fall into that category.

Interesting observation. A neural network is definitely a (complicated) formula, but anything computable can be written as a formula. So I think the statement that formulas are not copyrightable has to be qualified in some way (I don't know how).

[towforce]

A formula cannot be copyrighted - link - but I don't think that a set of NNUE weights would fall into that category.

Interesting observation. A neural network is definitely a (complicated) formula, but anything computable can be written as a formula. So I think the statement that formulas are not copyrightable has to be qualified in some way (I don't know how).

Something that's a mathematical truism like 2+2=4 cannot be copyrighted. However, an expression that gives rises to advice could be copyrighted: I think that the CANSLIM expression for stock selection would have been copyrightable when it was originally written in the 1950s.

Interestingly, this would mean that an EF which doesn't provide a perfect evaluation could be copyrightable, whereas an expression that gives a mathematically provable correct evaluation might not be!

[Ckappe]

As to the original question, "Are neural nets (the weights file) copyrightable?" that will depend on originality.
If, for instance, you simply launch a bevy of engines who labor for a few days and write the net, there is no originality in that.
If, on the other hand, you do lots of manual things (perhaps blending games from correspondence championships with games from TCEC and adding in specific test problems and endgame knowledge manually) then it might be copyrightable.

But I could be wrong about that. I think copyright issues are best resolved by people who know what they are doing, which means lawyers who have a degree in intellectual property laws

I guess that we have none of those in this forum/

I disagree with the premise in your first part wholly. Who is to say that hand selecting games is a more copyrightable process of building a network than the selection of the paramaters used to launch the fleet of engines?

This sort of vague idea that "Oh, my nets special because I want and hand picked out games from U13 mongolio", is the thing that Alberto Plata did at the start with DeusX; tricking the admins into thinking he had really gone above and beyond to build something special.

"First, before determining who may hold the copyright, we must first determine whether computer-generated art fulfills the basic requirements necessary to receive copyright protection. Copyright protection is available for 1) an original work of authorship, 2) fixed in a tangible medium 3) that has a minimal amount of creativity. If a work doesn’t have all three of these components, then it is not the copyrightable subject matter."

Machine-generation of a NN is hard to argue that a major effort of human creativity is needed to accomplish. The person arguing copyright must show that it required some personal/human creative unique effort for copyright to be possible, At least in US law.

[towforce]

"First, before determining who may hold the copyright, we must first determine whether computer-generated art fulfills the basic requirements necessary to receive copyright protection. Copyright protection is available for 1) an original work of authorship, 2) fixed in a tangible medium 3) that has a minimal amount of creativity. If a work doesn’t have all three of these components, then it is not the copyrightable subject matter."

Machine-generation of a NN is hard to argue that a major effort of human creativity is needed to accomplish. The person arguing copyright must show that it required some personal/human creative unique effort for copyright to be possible, At least in US law.

I've never heard of a case of computer generated art being refused copyright.

Music can be copyrighted, even if generated by electronic instruments.

[brianr]

Machine-generation of a NN is hard to argue that a major effort of human creativity is needed to accomplish. The person arguing copyright must show that it required some personal/human creative unique effort for copyright to be possible, At least in US law.

Creativity is certainly required to train some (but not all) nets.
I don't know if they are copyrightable.
Perhaps they patentable?
After all, just mixing a few chemicals may not be copyrightable, but the product produced seems to be patentable.
The algorithms and steps used to train a net may not by copyrightable, but the resulting net that plays chess might well be patentable.

[syzygy]

Again: think of the example of a GPL'ed game engine with copyrighted level files (I think most people would agree that level files are data). This scenario is very common (old versions of doom, quake,...).

Level files are works of art. They express something that is humanly perceivable and their creation involves an amount of creative freedom. The copyright on a JPG file is identical to the copyright on the image it encodes.

NN weights are not like that. At best they express a program's functionality, which is not copyrightable.

So if weights are data, there is no copyright on them.

The argument that NN weights are program code is interesting, but I don't think it works. They are not source code (whose creation involves some amount of free (non-functional) creativity, e.g. in naming variables, arranging the code, etc.) or drectly translated from source code.

[towforce]

Have Chessbase placed a copyright notice on their file of net weights?

If not, then IMO you could legally copy it.

If they have (I'm guessing that there's a copyright notice on the whole software bundle somewhere) then you cannot.

In practical terms, if you just made a copy for yourself, they probably won't pursue you (as long as you don't brag about what you've done in forums). If you make a copy available for others to download, then they probably would.

I think that it's very likely that a set of NN weights can be copyrighted, just like any other software can - even if it has been compiled, transpiled or whatever.

[Milos]

"First, before determining who may hold the copyright, we must first determine whether computer-generated art fulfills the basic requirements necessary to receive copyright protection. Copyright protection is available for 1) an original work of authorship, 2) fixed in a tangible medium 3) that has a minimal amount of creativity. If a work doesn’t have all three of these components, then it is not the copyrightable subject matter."

Machine-generation of a NN is hard to argue that a major effort of human creativity is needed to accomplish. The person arguing copyright must show that it required some personal/human creative unique effort for copyright to be possible, At least in US law.

I've never heard of a case of computer generated art being refused copyright.

Music can be copyrighted, even if generated by electronic instruments.

You are seriously mixing things.
Art generated using electronic tools/instruments is certainly copyrightable. "Art" generated exclusively by AI is not (quotes are there because this is not art in most common meaning of the word).

[Ozymandias]

NN weights are not like that. At best they express a program's functionality, which is not copyrightable.

So if weights are data, there is no copyright on them.

The legal question is outside my purview, but ethically speaking is very clear CB has had no problem producing data with freely available programs before, or taking it directly, and then charging a buck for them. In the first case they say you can't do whatever you want with that data because it's their property to do and sell as they see fit. In the second, they just try to hide the source of the data as much as they can, so that their buyers will be oblivious to the fact that they can get it for free. (Did you know you have to get all the way down to the Endgame Turbo 4 product description, to even know what's the actual name of that which you're buying, much less what it actually is or who made it?).