Copyright and Machine Learning IP

[syzygy]

Entering the Louvre and photographing the Mona Lisa is also forbidden. Therefore photos of Mona Lisa unless authorised would have been taken illegally and so on. Likewise interior of Cistine Chapel. Photography not allowed. You want sell pictures? Show contract that overruled the no photography rule.

To stop you, the Louvre or the Pope will at least have to prove that you were the one that broke the house rule, i.e. that was bound by a contract.

Likewise pop concerts, filming not allowed.

Here the artist has a stronger claim since he has a right on his performance. And the holder of the copyright on the music also has a claim obviously. Da Vinci and Michelangelo have no claim as they've been dead for too long.

[syzygy]

Note that he said the training set was a database, not the NN.

Ah, you're right. I completely misread the post I responded to.

Anyway, I think I have answered my own question earlier in this thread about whether an NN might be protected in the EU by a database right. (There might still be the argument that collecting the games qualifies as collecting the material in the NN because the NN weights are produced from the games, but I don't think this counts.)

[chrisw]

Note that he said the training set was a database, not the NN.

Ah, you're right. I completely misread the post I responded to.

Anyway, I think I have answered my own question earlier in this thread about whether an NN might be protected in the EU by a database right. (There might still be the argument that collecting the games qualifies as collecting the material in the NN because the NN weights are produced from the games, but I don't think this counts.)

Has the point been made that NN weights can't possibly form a "database" because a database consists of discrete independent pieces of data organised in some way, alphabetically or whatever, you can sort them according to whatever criteria, amongst a mass of other possible operations. Individual records can be pulled out, and the records do not affect each other. Because independent. Alice could sort it one way today and Bob could still use it tomorrow for whatever purpose.

Meanwhile NN weights are thoroughly and completely dependent on each other. Just try and perform a sort, please. And then see what you get. Database? No way.

[hgm]

I think no one here thinksis of the opinion that NN weights are a database. This was just a misunderstanding. So no need to flog a dead horse.

I think the structured set of weights constitutes the object code of a computer program, though.

[hgm]

Let me try the same logic on a computer with a (currently) more conventional computer architecture: "i86 code cannot be perceived by the human senses, so there is no copyright on computer programs". Doesn't sound right to me.

You can disassemble it and then it makes sense.

That is debatable. To most people it would not make sense at all. Even to programmers it would only start to make sense after very deep analysis. It would only be obvious to them what small snippets of code do, but that would not tell them what the role of that action is in the calculation.

NNs can also be disassembled, into a sequence of assignment statements of a conventional programming language.

I am also a bit uncomfortable with this focusing on EU law, now you mentioned that member states are free to adopt laws to protect sports events not protected by EU law. That woke me up to the fact that EU law is not a complete system of laws of a civilized society, but just the part of it that could be harmonized over all its member states. So it is possible that things not protected by EU laws on IP are in practice still protected IP in virtually all its member states, but through inhomogeneous nationla legislation. "Not forbidden by any EU law" might not mean all that much.

[Fulvio]

Note that he said the training set was a database, not the NN.

Ah, you're right. I completely misread the post I responded to.

To clarify, the point I wanted to emphasize is that the copyright of the data does not transfer to the NN. In the example of Bob Dylan's songs it is necessary to acquire the rights to use them, but the NN does not become a copyright of Bob Dylan.

But I did not mean to imply that there can never be a copyright / patent on an NN.
For example, in weather forecast, the software analyzes some data and creates animated images from them. There is no copyright on the forecast, but there is on the images because they present the data in an original way.
The same applies in my opinion to a high-quality list of pseudo-random numbers, there is no copyright on a single number, but the list represents an original work that is protected by copyright.

I do not think it is possible to say that an NN is never or is always covered by copyright or that is not patentable.
It is a complex subject with many variables. The NN may be created in a proprietary codec (like with video codecs for example), or the software that uses the NN may have a license that prohibits the use of unapproved NNs, the process of creating the NN may not be completely automated, etc ...

[syzygy]

But I did not mean to imply that there can never be a copyright / patent on an NN.
For example, in weather forecast, the software analyzes some data and creates animated images from them. There is no copyright on the forecast, but there is on the images because they present the data in an original way.

There is copyright on the graphical elements (if original) but there is no copyright on how they are arranged to present today's weather data.

The same applies in my opinion to a high-quality list of pseudo-random numbers, there is no copyright on a single number, but the list represents an original work that is protected by copyright.

By definition there can be nothing creative in a high-quality list of pseudo-random numbers. (A non-random list of numbers will not be copyrighted either, unless they somehow encode/represent a copyrighted work.)

I do not think it is possible to say that an NN is never or is always covered by copyright or that is not patentable.

In my view there are strong arguments why an NN cannot be covered by copyright even if a human comes up with the weights.

[syzygy]

Meanwhile NN weights are thoroughly and completely dependent on each other. Just try and perform a sort, please. And then see what you get. Database? No way.

If you just dump the weights as numbers in a database (forgetting about any structure you might lose by doing that) and that database is then protected, then that would be sufficient to get the NN protected as a collection of weights. And of course you could preserve any structure you want by encoding the weights as records that remember the position of the weight in the NN.

(But the resulting database would not be protected either by copyright or by a sui generis database right as far as I understand.)

[chrisw]

I think no one here thinksis of the opinion that NN weights are a database. This was just a misunderstanding. So no need to flog a dead horse.

I think the structured set of weights constitutes the object code of a computer program, though.

Well, the architecture of a neural net maps directly to a computer program. You could say that the net inputs (chess position eg) are variables, the fixed weights as program constants and the net architecture as the program. If you really wanted, you could unroll the program loops and insert the constant weights appropriately in each unrolled section. And represent the whole thing (weights and architecture) as source code. You could get the training program to spew the whole thing out as source or object code if you really wanted, at which point you really have the ingenious machine that really has written a useful program. It's written all the source code of an evaluation function for chess positions.

if we wanted to get really clever, we could program our ingenious machine to modify the net architecture (smartly choosing new architectures) and test them in a MCTS model chess match environment. Then our ingenious machine not only writes evaluation function software, but tests it, and spews out a never ending stream of better and better evaluators.

I await being informed this is purely functional and no copyright applies anywhere. Especially not to the original ingenious machime programmers/designers. Except to H.G.Muller who will own copyright because he smartly provided some test positions and pressed GO.

[chrisw]

Meanwhile NN weights are thoroughly and completely dependent on each other. Just try and perform a sort, please. And then see what you get. Database? No way.

If you just dump the weights as numbers in a database (forgetting about any structure you might lose by doing that) and that database is then protected, then that would be sufficient to get the NN protected as a collection of weights. And of course you could preserve any structure you want by encoding the weights as records that remember the position of the weight in the NN.

(But the resulting database would not be protected either by copyright or by a sui generis database right as far as I understand.)

then your "database" would not be a weights file alone but would contain all the necessary information about the net architecture as well (which could be stacks of layers) and therefore more than the actual weights file.

So, no, a weights file is not a database. It expects the decoding software to make sense of it, where the decoding software contains the architecture knowledge.