TalkChess.com

All numbers inside mp3 file are also no more than coefficients in a giant formula which describes a waveform. But these coefficients are copyrightable with no doubt.

whereagles wrote: ↑Fri Aug 10, 2018 6:57 pm milos, you might think tcec is a joke but fact is everybody is having great fun there.

Milos may need some help with that strange word you used.

Here ya go, dude! https://www.merriam-webster.com/dictionary/fun

WinPooh wrote: ↑Sat Aug 11, 2018 12:30 am All numbers inside mp3 file are also no more than coefficients in a giant formula which describes a waveform. But these coefficients are copyrightable with no doubt.

Mp3 file is a digital representation of the music. Copyright is on music independent of media format.
Weights in NN are not digital representation of anything, they are just numbers without underlying content.
This distinction might be too fine for some...

You can't patent math either, for obvious reasons.

Further, from https://en.wikipedia.org/wiki/Patentability, we have this:

"The patent laws usually require that, for an invention to be patentable, it must be:

Patentable subject matter, i.e., a kind of subject-matter eligible for patent protection
Novel (i.e. at least some aspect of it must be new)
Non-obvious (in United States patent law) or involve an inventive step (in European patent law)
Useful (in U.S. patent law) or be susceptible of industrial application (in European patent law[1])

Usually the term "patentability" only refers to "substantive" conditions, and does not refer to formal conditions such as the "sufficiency of disclosure", the "unity of invention" or the "best mode requirement".

Judging patentability is one aspect of the official examination of a patent application performed by a patent examiner and may be tested in post-grant patent litigation.

Prior to filing a patent application, inventors sometimes obtain a patentability opinion from a patent agent or patent attorney regarding whether an invention satisfies the substantive conditions of patentability."

Robert Pope wrote: ↑Fri Aug 10, 2018 11:34 pm

syzygy wrote: ↑Fri Aug 10, 2018 10:44 pm An interesting question is whether there is copyright on the neural net. In my view there is not, since no creativity went into creating it. The only possible creativity would be in selecting the games to feed it, but this selection was made on the basis of functional criteria.

I think there is definite creativity in creating it, as there are many training parameters that will affect the final output, as well as the choice of games. There's not a single obvious correct procedure to follow. I think Albert said it took 7 attempts on the same data to get something out that played good chess.

But all those attempts were attempts to get the network to learn to play good chess, not attempts to express his creative freedom.

(And even if someone "creatively" selected tens of thousands of games to train the network on, for example by selecting them on the basis of a subjective feeling of beauty rather than white's and black's Elo ratings, there would still seem to be a rather big question as to whether any of that creativity survives in the end result.)

Milos wrote: ↑Sat Aug 11, 2018 1:16 am Mp3 file is a digital representation of the music. Copyright is on music independent of media format.
Weights in NN are not digital representation of anything, they are just numbers without underlying content.

They are representation of chess knowledge.
Also, if I move these numbers inside C++ file, compile it and make initial NN built-in in the program's binary, will they become copyrightable/patentable?

Neural nets are just a method to encode knowledge. There is non-trivial work involved in creating (designing the topology and training) them. Surely the result counts as intellectual property. Even when I publish a book to present a collection of chess games (which in themselves are not copyrightable), the book will be protected by copyright. People would be allowed to copy and spread the individual games, but not the collection, which is 'tainted' by my efforts to select the games.

hgm wrote: ↑Sat Aug 11, 2018 1:05 pm Neural nets are just a method to encode knowledge. There is non-trivial work involved in creating (designing the topology and training) them. Surely the result counts as intellectual property. Even when I publish a book to present a collection of chess games (which in themselves are not copyrightable), the book will be protected by copyright. People would be allowed to copy and spread the individual games, but not the collection, which is 'tainted' by my efforts to select the games.

Similar story with tsumego (tactical go problems) in Japan. Individual tasks are not copyrighted, collections without solutions are not copyrighted. But complete books of problems are protected by law.

hgm wrote: ↑Sat Aug 11, 2018 1:05 pm Neural nets are just a method to encode knowledge. There is non-trivial work involved in creating (designing the topology and training) them. Surely the result counts as intellectual property. Even when I publish a book to present a collection of chess games (which in themselves are not copyrightable), the book will be protected by copyright. People would be allowed to copy and spread the individual games, but not the collection, which is 'tainted' by my efforts to select the games.

As I understand it, the trained neural net weights (being specific here about the weights used in Leela Chess) are produced by applying a large data set of chess games to a Leela Chess Training Machine, where we can say "abstract unknown chess knowledge" has been encoded or abstracted somehow into the weights, as part of a process where the creatively encoded/abstracted "chess knowledge" extracted from the data set of games came from the creative actions of the Leela Chess Training Machine.

To meaningfully use these weights requires them to be used an inputs to a Leela Chess Playing Machine.
There's an interest in copyright, I assume, because someone else other than Leela Authors has put another large data set through the same Leela Chess Training Machine and produced another set of weights. I read that the selection of these games is functional rather than creative.

For copyright to apply, we need a) human(s) (machines can't own copyright) and b) creative acts to have been carried out.

Two possibilities arise.
1. Nobody owns copyright in the weights because they are machine generated, where the operator provided no creative input, other than press GO.
2. Leela Authors own copyright because they wrote the algorithms that creatively abstracted/encoded the chess knowledge.

From what I read it is probable that legal systems, which are still catching up with ML consequences, will apply option 2.
Note that in no case does the operator who put the other set of games through Leela Chess Trainer get any copyright at all.

The consequences of this, either way, 1 or 2, is that the entity "DeusX" is the intellectual property of Leela Authors. And therefore, whether it plays, doesn't play, has parameters set or not, is not a matter between anybody else other than Leela Chess Authors and TCEC.

It is debatable whether there is any creativity involved in training a NN. It is just mechanically applying a rather simple mathematical process in bulk. IMO the optimizer of a compiler is more creative. Yet everyone would agree that GNU doesn't own any copyrights on executables produced by gcc.

My take is that if there was no creative effort in selecting the training games, then the DeusX weights are public domain. Not that it matters a hoot; even if they were copyrighted, anyone could get the same weights by non-creatively feeding the games to LC0 and pressing GO, and those weights would then be free for them to distribute. Because they were not copied, but independently generated.

The LC0 authors have parted with any hypothetical rights they might have been allowed to claim on whatever is produced through their code, by releasing that code under the GPL. The only thing they can require is that the code itself is accompanied by the source. It is not clear whether the NN weights should be considered code or data.But even if they are code, anyone that distributes them only would have to accompany them by their 'source', it is the collection of training games, in order to make it legal.