syzygy wrote: ↑
Sun Aug 12, 2018 7:25 pm
chrisw wrote: ↑
Sun Aug 12, 2018 2:51 pm
An MP3 file consists of a 1-D structured set of numbers. But it represents music. That's copyrightable because of what it represents. Helpfully it's a one-to-one mapping, and decodable by MP3 player. And the music is tangible to the ear.
This is a good example.
If you feed an mp3 encoder copyrighted music, the resulting mp3 file will be covered by the copyright on the music (because the mp3 file easily preserves enough of the human-perceivable features of the music). It will not be covered by the copyright on the mp3 encoder because the mp3 encoder determines only the functional structure of the mp3 file.
Instead of an mp3 encoder taking as input music and producing as output an mp3 file, here we have an NN trainer program taking as input a collection of games and producing as output a set of weights. For the set of weights to be copyrighted we at least need a (somewhat) originally collection of games and to somehow be able to convince a judge that enough of that originality is perceivable from the set of weights. That's going to be tough.
At the moment a case is pending before the CJEU in which the Court will have to decide whether a taste can be a copyrighted work. The advocate general recently issued his opinion advising the court to decide that the notion of "work" encompasses only subject matter that can be perceived through sight or hearing and "with precision, stability and objectivity".
A neural network as a set of numbers can be perceived by the human eye but is then just meaningless. One will not be able to say that a particular network infringes on the copyright of another network by just looking at the similarity between the weights. (The same holds true for an mp3 file: as a set of numbers it is meaningless to the human eye or ear. You have to run it through an mp3 player to perceive it meaningfully.)
So a neural network will have to be run through the Lc0 client, but can we then perceive it "with precision, stability and objectivity"? I'm not convinced that "playing ability/style" is something that can be the object of copyright. Of course playing strength can be measured to some extent, but that is just a functional criterion (like how aerodynamically a car is shaped).
It is possible that the criteria formulated by the CJEU will be more relaxed than those proposed by the advocate general.
The NN represents chess knowledge. It's not any form of one-to-one mapping, it's intangible, and it's decodable by LC0.EXE and into LC0.EXE language and performing weird transformations on it and then outputting chess moves AND THEN a smart chess knowledgable person decoding what she thinks the system "knows". OR, by a possibly less smart person comparatively looking at the numeric move value outputs.
..... the NN as a structured set of weights. There is no copyright on that structure.
Because the LCO(Trainer) carries out a repetitive, functional series of steps on the data. Not creative. Therefore no copyright. I think is your argument.
My argument is that the Lc0 code does not determine the values of weights but probably only the number of weights. If Lc0 requires that the NN has 5x5 weights, then you will agree that the copyright on Lc0 does not extend to all collections of 5x5 numbers. Even if it determines much more structure than something like "5x5", it will not be enough because the structure is just that what is necessary to make it work with the Lc0 client. Just like the structure imposed by an mp3 encoder is just that what is necessary to allow the mp3 file to be decoded by an mp3 player. (There is probably a bit more to it in case of mp3 encoders since they do not all produce the same quality output, but any variation between mp3 encoders is not an expression of creative freedom but of an attempt to reach a high quality or a good quality/time trade off -- technical criteria. And the perceivable differences between two encodings of the same music or between an mp3 file and the original music will anyway be extremely unlikely to establish a new/additional copyright.)
On Advocate Generals ... well, yes, but. Humans are used to copyright applying to tangible, maximum 3-D things. We are entering a world of intangibles and many dimensions. Maybe super-smart stuff that we can't understand just isn't amenable to the ancient law of copyright.
On weight numbers ...
I think LC0.EXE is able to use weights from past periods of training when different net sizes were used. Probably there's some code to detect net type-structure and LC0.EXE adjusts its software accordingly.
On your "perception" points ...
Because the net is modelled to some extent on (human) visual system, it might be possible to detect visual patterns forming in the net layer neurons as the "understanding" becomes deeper, so to speak. I would not like to be the poor guy given the task of hunting for these across different game positions inputs, comparing them, and trying to generate some visual material to convince your advocate general, though.
Nevertheless, weights do apparently show distinct patterning themselves, viewed as layered heat maps, so perhaps detecting if one net is a retrained derivative of an earlier net might also be possible. Again, I would not want to be the person given the task of hunting. In all cases we would be suffering from the how close is close problem and how in general to test for nets trained or more or fewer or different games, how similar they are anyway.
Probably the best way to "protect" weights (commercially) is to protect the reading code, since that's a vital component of using the weights.
Okay, I am outnumbered by good arguments. I drop that machine generated weights can be copyrighted. Does that apply to PST's too? Runs and hides ....