Hacker News Reader

Comment by inlined | Hacker News Reader

Forgive my ignorance of ML but the last bit: "you'll need your own porn to train on" confused me. Does this mean that they're just exposing the rough topology of their neutral net (eg depth) and not the actual weights between nodes? I'm curious to learn from an ML expert how much this actually offers.

View on HN

It looks like they are sharing the trained network, but they aren't sharing the training data set.

The training set is almost certainly composed of copyrighted material.

Interesting thought - doesn't every single porn producer now have a valid copyright claim on the trained network? I don't see how you can argue this isn't a derivative work based on the movies they produced.

I was debating with a friend about just this: whether a text-to-speech model is a derivative work of audio recordings by a given speaker, such that they'd then have claim on ownership of it. (You could almost certainly create an [overfit] model that could re-generate the original performance of a text from said text.)

Moot if it was a work-for-hire, of course; but if I, say, created a Samuel L. Jackson speech model by training on samples from his movies, and sold it as one of those car-navigation voices, could I be sued? By Mr. Jackson? By the copyright-holders of the movies?

And if I could, what does that imply about impersonators, who do the same thing, but with their brains?

I imagine you couldn't put his name on it which would be a huge deal if you wanted to sell it.

I don't think it's a derivative work just because one of the inputs is copyrighted. I think it's more descriptive than derivative. Content producers don't generally own copyright in critics' descriptions of their movies or of plot summaries, even though their copyrighted material is a necessary input to the description's creation.

Also, good luck trying to prove that your film was used to train this network.

You could be compelled to produce your training set during the discovery phase of a lawsuit.

Only if there was any reasonable basis to believe that infringement had taken place. Most countries have some sort of pre-trial hearing before a civil suit to determine if there's merit to the accusation. You're generally not allowed to go on fishing expeditions without a reasonable basis for your claim(s).

It would not be considered derivative work, because what is produced is nothing like the original. There is nothing recognizable in the work produced for a court to rule on.

This is like saying the hash of the text of a book is derivative. If it were ruled that this is the case (that a hash is a derivative work) then suddenly every single number in existence is a derivative of every single other number (since there will always exist some function that will transform X into Y.)

It would fall under "transformative" fair use.

Except the 'derivative' has nothing in common with the original.

I don't think you could argue this is a derivative work and not be sanctioned by the courts for bringing frivolous litigation.

If you want to train yes. AFAIK, they are releasing a pre-trained model which you can just use right away. There is no sense in "topology" of this yahoo specific model as it is a caffenet.

You initialized the nets using their weights, and then provide your own data, in this case, a list of images contains a (porn|no-porn) label to 'fine-tune' the nets towards your usage case.

Reply on news.ycombinator.com