Stockfish: Our lawsuit against ChessBase

[syzygy]

If you manage to use NNs as a functional alternative to JPEG, then sure, an NN encoding a specific image will be covered by the copyright on the image. If you manage to use an NN to encode the source code of a computer program, then same.

Well, that confirms the only point I wanted to make: the reasoning"these are weights of an NN, so they cannot be protected by copyright" is not sound. It depends on what the NN does, and how it was taught or designed to do that.

I expressed myself more carefully than that:

The NN weights are not copyrighted. Just numbers not encoding any expression.

When in some exceptional case they do encode expression, then sure they are covered by that copyright.

I have never seen an NN that encodes a specific image. I wouldn't call it an NN. But an image created by a human will usually be copyrighted.

Now in the case of Chess we have of course to realize that there is a court ruling that declares the moves of a chess game cannot be copyrighted. This makes it less likely that even the data fed to the NN for training could carry any copyrights that the NN might inherit (as a derived work).

Fisher's 60 memorable games are probably a creative selection that is copyrighted. Fisher selected them based on personal, subjective criteria, and clearly this selection reflects his personality and bears his personal stamp, etc.

The selection of all games played by GMs with > 2600 Elo, for example, would not be a creative selection. Any selection intended to result in an NN that plays strong chess is functional and not creative.

Even if the selection of training material is copyrighted, I very much doubt that it survives in the NN. But one could try to convince the judge. (Most of the expression in source code does not survive in the object code, and it is probably accepted that the copyright extends to the object code. But on that we have relatively clear legislative intent, and object code is very close to source code and arguably still has some expression. It would be interesting to have a decision by SCOTUS or the CJEU analysing this in detail.)

[towforce]

This issue simply cannot be unique to NNs.

From what I'm reading here, if a company sells machines with no patents in a pretty container, I cannot copy the container because that has "artistic human expression", but I can copy the machine EXACTLY.

If that was the law, then I would have expected it to have happened many times. I know there are allegations of developing countries doing this in the pre-internet era (I've actually seen this done to me: I wrote an article for a magazine, and years later saw my article, with very few changes, in a Brazilian magazine - credited to somebody I'd never heard of!), but that's not what I'm taking about: here's what I'm talking about:

* a factory near to where I live produces a useful machine in a pretty container, and it has no patents

* I commission a different factory to make the EXACT SAME MACHINE (no differences at all) and package it in a different container

If there's no legal reason for me to not do that, why doesn't it happen all the time?

I guess cos nobody has interest in an steam machine or an decades old CPU design, patents are usually 20 years, mask work 10 to 15, Copyright somewhere between 50 to 100, you know probable the .mp3 and .jpeg patents (technical solution), Frauenhofer Institute collected millions in fees with .mp3, nowadays everybody is free to implement encoder/decoder without fees, but meanwhile new codecs offer new features, project https://www.gutenberg.org/ has thousands of free books online, and there were legal issues with this cos countries differ when Copyright does expire...the Copyright on the original Mickey Mouse character expired recently, but Disney changes the design over the years, even if you are now free to use the original Mickey, nowadays kids won't recognize him...

---
Srdja

Very interesting - thank you. However, Mickey Mouse is a work of art - not a machine.

Why don't businesses exist that do nothing more than make EXACT copies of new machines that don't have patents on them and put them in a new outer layer?

Or maybe the remedy to free-riders copying NNs (apart from having an app which converts NN weight files to lines of code, which get protection) is an app to apply for a patent for an NN, making it a quick and easy thing to do?

Here's a simple, unmistakable expression of what I'm saying: "you can't just copy other people's products and sell them as your own".

This is a general rule, of course: there may be exceptions.

[smatovic]

[...]
Here's a simple, unmistakable expression of what I'm saying: "you can't just copy other people's products and sell them as your own".
[...]

Not sure what you are up to (as metaphor for neural networks), but did you ever hear from the "IBM PC compatible computers"?

https://en.wikipedia.org/wiki/IBM_PC_compatible

IBM did not patent the microarchitecture (unlike its PS/2 Micro-Channel from 1987), Intel and AMD did deliver CPUs, the BIOS was clean-room-reverse-engineered, MS-DOS was indepented -> IBM compatible PC.

The dark side of an open system is its imitators. If the specs are clear enough for you to design peripherals, they are clear enough for you to design imitations. Apple ... has patents on two important components of its systems ... IBM, which reportedly has no special patents on the PC, is even more vulnerable. Numerous PC-compatible machines—the grapevine says 60 or more—have begun to appear in the marketplace.

--
Srdja

[smatovic]

If you manage to use NNs as a functional alternative to JPEG, then sure, an NN encoding a specific image will be covered by the copyright on the image. If you manage to use an NN to encode the source code of a computer program, then same.

Well, that confirms the only point I wanted to make: the reasoning"these are weights of an NN, so they cannot be protected by copyright" is not sound. It depends on what the NN does, and how it was taught or designed to do that.

I expressed myself more carefully than that:

The NN weights are not copyrighted. Just numbers not encoding any expression.

When in some exceptional case they do encode expression, then sure they are covered by that copyright.

I have never seen an NN that encodes a specific image. I wouldn't call it an NN. But an image created by a human will usually be copyrighted.

Now in the case of Chess we have of course to realize that there is a court ruling that declares the moves of a chess game cannot be copyrighted. This makes it less likely that even the data fed to the NN for training could carry any copyrights that the NN might inherit (as a derived work).

Fisher's 60 memorable games are probably a creative selection that is copyrighted. Fisher selected them based on personal, subjective criteria, and clearly this selection reflects his personality and bears his personal stamp, etc.

The selection of all games played by GMs with > 2600 Elo, for example, would not be a creative selection. Any selection intended to result in an NN that plays strong chess is functional and not creative.

Even if the selection of training material is copyrighted, I very much doubt that it survives in the NN. But one could try to convince the judge. (Most of the expression in source code does not survive in the object code, and it is probably accepted that the copyright extends to the object code. But on that we have relatively clear legislative intent, and object code is very close to source code and arguably still has some expression. It would be interesting to have a decision by SCOTUS or the CJEU analysing this in detail.)

Just one more note, usually in chess they use labeled positions/games for training NNs, for example a position score from an engine with hand-crafted evaluation function, hence you train the NN against an copyrighted evaluation function in such a case.

--
Srdja

[syzygy]

I think the line draws similar to a dictionary, like you can't copyright individual words, their meanings or explanations, however you can copyright the specific instance of their collections, which involves a fair amount of human involvement.

The question is not whether enough time was invested, but whether the collection is a creative expression.
forum3/viewtopic.php?p=940194#p940194

An alphabetic list of the words of a language that the author of the list tried to make as complete as possible is not protected by copyright. A normal dictionary contains more than just a list of words and will normally be copyrighted, but if you copy the list of words into a file to make a spell checker, you should be fine. (There is a Dutch case where the appeal court finally decided that drawing up the list of words inevitably involved subjective decisions and that the list was therefore copyrighted, but it seems to me that a good dictionary will simply try to be complete and free of errors. But evidently there is room for discussion when it comes to drawing up a list of words for a dictionary.)

The numbers that make up an NN, even if hand picked (which they are not), are intended to provide functionality. Functionality is not copyrighted. There is no other idea behind the numbers. (For HGM: sure, if the functionality involves displaying a specific image it will be different. The creator of that image - not the creator of the NN weights - might have a copyright.)

"[C]opyright is limited to those aspects of the work -- termed 'expression' -- that display the stamp of the author's originality" (Harper & Row v. Nation Enterprises). NN weights do not display the stamp of any human's originality.

It also does not contradict the steam engine example, the idea can't be copyrighted, the specific designs can be covered by patents, the specific engine(instance) can be copyrighted.

There is no such thing as copyright on a steam engine, unless you embellish it with non-functional engravings etc.

Copyright only asks for a modicum of creativity, but "hello world" is too banal for that.

So why is there a problem?

For the purpose of this thread, I don't care what anyone would like the law to be. I care about what the law is.

[noobpwnftw]

What the law "is" however is a subjective matter, except for the very far fetched ones, that are often decided by things other than what's written in a textbook.

[hgm]

I have never seen an NN that encodes a specific image. I wouldn't call it an NN. But an image created by a human will usually be copyrighted.

These were already known in the very early days on NN research. You just take a large grid of cells which act as a retina, excited by the corresponding pixel of an image, as well as output, their activity controlling a pixel in a display. You make a large number of random bidirectional connections between the cells. The learning rule is that connections between cells that are active (= excited above their threshold) simultaneously get their weights increased, while connections where only one end is active get their weight decreased.

You then train the net by showing it a set of images again and again. After some time the NN will be able to recall each of the training images entirely (i.e. show it in the diaplsy) by feeding it a unique part of such an image as input. And will then retain the recalled image even after you remove the input stimulus.

[syzygy]

the Copyright on the original Mickey Mouse character expired recently, but Disney changes the design over the years, even if you are now free to use the original Mickey, nowadays kids won't recognize him...

It seems the original copyright will expire in the US exactly 1 year from today.

In Europe it should expire after 70 years counting from the first day of the year following the death of the author. Walt Disney died in 1966, which means his copyrights should expire on January 1, 2037. (But if Disney the company is the "author", then it would expire after 70 years counting from the first day of the year following the first publication. It seems Mickey Mouse was first published in 1928, so that would have been 1 January 1999. I suppose Walt Disney the person is the author, though.)

[syzygy]

I have never seen an NN that encodes a specific image. I wouldn't call it an NN. But an image created by a human will usually be copyrighted.

These were already known in the very early days on NN research. You just take a large grid of cells wich act as a retina, excited by the corresponding pixel of an image, as well as output, their activity controlling a pixel in a display. You make a large number of random bidirectional connections between the cells. The learning rule is that connections between cells that are active (= excited above their threshold) simultaneously get their weights increased, while connections where only one end is active get their weight decreased.

You then train the net by showing it a set of images again and again. After some time the NN will be able to recall each of the training images entirely by feeding it a unique part of such an image. And will then retain the recalled image even after you remove the input stimulus.

OK, such NNs would seem to fall under the copyright of the images. This means that not whoever is responsible for the weights has a copyright, but whoever owns the copyrights on the images.

[syzygy]

What the law "is" however is a subjective matter, except for the very far fetched ones, that are often decided by things other than what's written in a textbook.

The law is meant to be objective, at least wherever the rule of law is a thing. There are always grey areas, and applying the law to the facts of a case may involve making judgments with a subjective component, but it is certainly possible to discuss what the law "is".