Are neural nets (the weights file) copyrightable?

[towforce]

Case: ING, a Dutch company, commissioned a "new Rembrandt" to be generated by AI and printed on a 3D printer. There hasn't been a legal test case, but the article below argues that the copyright for this new painting probably belongs to ING:

https://alj.artrepreneur.com/the-next-r ... rated-art/

On this basis, assuming that Chessbase commissioned Mr Silver to produce the FF2 NN, and further assuming that a court would view this NN as an original work (hmm........), then the copyright for that NN would belong to Chessbase.

[chrisw]

The general claim that NN weights are not copyrightable seems untenable. Such weights could in principle be hand-crafted, the result of a creative design process, written as source code and compiled into a binary file of weights. Any network of boolean logic can be emulated by a neural net. Conventional hand-crafted evaluations are in fact neural very simple neural networks. The binary resulting from the compilation of their source code carries copyrights.

Nonsense. NN weights are streams of numbers and streams of numbers (pi, e and so on) are not copyrightable unless they are demonstrably special in some way. Simply claiming copyright is insufficient, you would have to show that the number stream represented something that was created by human hand. A song, a book, some computer code, a database, whatever. Obviously you can’t put the number stream directly along side the “original human created work”, so, ultimately, you would need to perform AFC test, abstract out of the number stream something that could be put side by side with whatever is claimed as the human created copyrighted thing and then claim a copy-match.
But the NN can’t be reversed backwards, so you can’t do AFC, so any copyright claim is dead in the water.

[dkappe]

Copyright law is complex, and us non-lawyers can have opinions about it, but what is copyrightable is not always clear-cut, doesn't always follow what seems intuitive, and can vary by jurisdiction.

Exactly. As mentioned at the very start of this thread, I have integrated licensed pretrained models into a commercial application on behalf of of a client. The client’s IP lawyers gave a typical lawyerly opinion on whether pretrained models could be copyrighted and licensed: probably, but it has yet to be tested in court.

In the meantime, here are a sampling of projects and companies claiming licenses (which generally presumes copyright) on their pretrained models:
- pytorch project
- ImageNet (to the point where models trained on ImageNet would be covered)
- Facebook
- Google

There is also an AWS marketplace for pretrained models as a service. (https://aws.amazon.com/marketplace/solu ... ned-models)

Last, pretrained models have been in use in medical imaging solutions for years and are generally covered under the commercial license of the product.

My money is on the courts and Congress in the US siding with commercial interests.

[towforce]

Nonsense. NN weights are streams of numbers and streams of numbers (pi, e and so on) are not copyrightable unless they are demonstrably special in some way. Simply claiming copyright is insufficient, you would have to show that the number stream represented something that was created by human hand. A song, a book, some computer code, a database, whatever. Obviously you can’t put the number stream directly along side the “original human created work”, so, ultimately, you would need to perform AFC test, abstract out of the number stream something that could be put side by side with whatever is claimed as the human created copyrighted thing and then claim a copy-match.
But the NN can’t be reversed backwards, so you can’t do AFC, so any copyright claim is dead in the water.

The AFC test (link) could readily be applied to a net. One of the following two conditions would do:

1. the weights files being compared were very similar

2. the two programs gave very similar evaluations (without branching) in a wide variety of different positions (this wouldn't work if the evaluations were actually 100% chess-accurate)

[Ferdy]

the protection with regards to ai there may not be underestimated.

Interesting. Just the act of monitoring a process and stopping when the results are artistically pleasing is probably enough, while a pure supervised learning would not be. I am confused why they say a reinforcement learning also counts. Presumably there is a human directing some of the reinforcement direction or at least monitoring hyperparmeters and changing them and determining when to stop the process? That could apply to supervised learning too, so not sure what the distinction they are making is.

It is the unsupervised learning that has the low or zero probability of being granted a copyright. Reinforcement learning in autonomous cars is not easy, the ai systems, NN weights, etc. there has a high probability of getting a protection.

Things may favors NN in the long run, as it is also about the cost of producing the NN weights that also matters. People, Corporations have invested a lot on ai, what is the point of this investment if the NN weight is not protected. The common phrase "only works from human being can be protected under Copyright Act." probably needs reexamination. Programmers can make ai system automate many things a lot more if there are more resources. Are we going to hold back the ability of programmers and not fully use the resources available by letting the human do some works instead of ai so that our NN weights may get a high chance of getting a protection?

[Ckappe]

The most sensible way will be to look at NNs as a sub-routine to the code using the weights.. so the NN is seen as an integral part of the original program using it, not as a separate stand-alone entity. and thereby it would always fall under and inherit the copyright, licensing of the main().
Just as a few variable names is not copyrightable by themselves the sum of code including these names and weights could be copyrighted as a whole.


This will also solve most commercial aspects. It would not grant any copyright to stand-alone "NETs" but it would protect AI systems where the NN is a part to produce a complete AI.

My view is that current old IP-rights legislation, in general, is broken in this digital era. It's a huge problem and blocker of innovation by itself today by focusing more on the "ownership", IP-holder rights, than the IP-creator, prior art, uniqueness, the common good, and the promotion of knowledge sharing.

[towforce]

The most sensible way will be to look at NNs as a sub-routine to the code using the weights.. so the NN is seen as an integral part of the original program using it, not as a separate stand-alone entity. and thereby it would always fall under and inherit the copyright, licensing of the main().
Just as a few variable names is not copyrightable by themselves the sum of code including these names and weights could be copyrighted as a whole.

I agree. I don't know, but my opinion is that if you took GPL code and moved some classes into a separate library, but still used them in exactly the same way, you'd still be obliged to GPL the new code. If this is correct, then it strongly implies that FF2's NN weights file should be under a GPL license.

[dkappe]

I agree. I don't know, but my opinion is that if you took GPL code and moved some classes into a separate library, but still used them in exactly the same way, you'd still be obliged to GPL the new code. If this is correct, then it strongly implies that FF2's NN weights file should be under a GPL license.

Don’t make me quote the GPL FAQ again. In short, no.

[gaard]

The following thought occurred to me... if Chessbase is claiming that the FF2 weights file is not subject to the GPLv3 that is tantamount to saying that the weights file is not subject to copyright.

I don't think that follows. It could be that they believe it falls under a different license.

As to the original question of if they are copyrightable at all, FWIW, many big businesses seem to think so and have gone to great lengths to detect infringements for their own benefit and that of their clients. For example, see https://www.ibm.com/blogs/research/2018 ... ermarking/

[Michel]

I agree. I don't know, but my opinion is that if you took GPL code and moved some classes into a separate library, but still used them in exactly the same way, you'd still be obliged to GPL the new code. If this is correct, then it strongly implies that FF2's NN weights file should be under a GPL license.

Don’t make me quote the GPL FAQ again. In short, no.

Can you please stop this. It has been amply demonstrated that the GPL faq does not cover the SF situation where the interpreted program enhances the interpreter.