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.
Tainted not so much by your efforts as by the creativity (if any) involved in making the selection.
If you select the games using objective criteria like minimum Elo rating of the players or level of the tournament, there will not be any copyright on the selection even if it cost you blood, toil, tears and sweat to collect them.
The topology of a neural net is chosen based on objective/functional criteria and will therefore not be protected by copyright. Training the net by burning a lot of electricity does not require creativity. Did creativity go into selecting the training games? I highly doubt it. (We seem to agree, judging by your other posts.)
hgm wrote: ↑Sat Aug 11, 2018 5:43 pm
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.
It's just data in my view. In any event there is no source code behind the weights that one could ask to be released, unless one considers that the collection of training games is the source code and the training program the compiler (which I would not subscribe to).
Well, as I said, I am not sure about this. But it is an interpretation I would not discard off-hand. A NN is an alternative architecture for a programmable data-processing machine, an it is programmed by setting the weights. The machine usually is virtualized, by emulating it on a conventional computer or video card.
This is not very different from the situation we have with an interpreted language like Java. For web applications this is usually compiled to (platform-independent) byte code, which acts as a program for a virtual Java machine, implemented through a (platform-dependent) interpreter in the browser. Now is the byte code program or data? As is usual in such a layered architecture, it depends on which level you are talking about. For the interpreter the byte code is just data. But for the virtual machine it is a program.
