The real 'knowledge' or the learning that makes a neural network give results is the values behind of the edge weights which are literal expressions (numbers) as seen below. The neural network is essentially useless without these appropriate values and hence this is an extremely valuable intellectual property. Is copy-right protection available for this? What all elements of a neural network get IP protection?

enter image description here

  • Well, patenting them would make them public and it could be hard to notice any infringement. And then, small changes to the numbers might fall outside of the protection. Is there a way to determine the exact ranges for the weights that yield good results? Then patenting might be an option, else, how would you even prove someone took your weights if they changed them a bit? And is keeping them secret not an option?
    – user18033
    Commented Jan 13, 2018 at 12:09
  • Neural networks have been known for a while now. I sincerely doubt you can get a patent on them even for a specific set of values. Where exactly is the inventive step?
    – Eric S
    Commented Jan 13, 2018 at 15:05
  • @EricShain the specific values are not obvious ;) I don't know if it will work, but it could.
    – user18033
    Commented Jan 13, 2018 at 15:43
  • @DonQuiKong abstract math isn’t patentable either. Non obvious does not equal novel.
    – Eric S
    Commented Jan 13, 2018 at 15:50
  • @EricShain novelty won't be the problem. I don't know, but maybe someone has an answer
    – user18033
    Commented Jan 13, 2018 at 16:10

2 Answers 2


Do the weights of a neural network get any kind of IP protection?

Almost certainly not, as they are obvious (in a patent sense). The reasoning for this is a little bit indirect.

Neural networks in general are very well-known and commonplace. In addition, training methods for neural networks are very well-known. By applying a known training method to a known neural network, you are doing nothing more than what has been done thousands of times before. The fact that you have a unique series of values as an output is not in itself non-obvious, since they were obtained by well-known and obvious techniques. As such, it would be obvious for the person skilled in the art to try the same approach, and they would then obtain substantially the same results.

Unless you have invented a unique type of neural network, or a unique method for training, or a unique step of filtering data, or the neural network is being applied in a suprising area, then you're really doing nothing non-obvious.


Copyright is off-topic here, and I would encourage you to post a similar question on Law.SE to get a full answer.

However, as an aside, almost certainly there would be no copyright in the weights of a neural network, since they would not involve even a modicum of creativity.

  • 1
    "As such, it would be obvious for the person skilled in the art to try the same approach, and they would then obtain substantially the same results." But would they? I'd argue that selecting a favorable option from a infinitude of options is inventive. That the well known method was applied in a well known way to a very well (maybe luckily) selcted training data set, making the output unexpectedly well suited for the task. And the burden of proving me wrong would (in theory) lie with the examiner. I don't know if the argumentation would lead to a patent, but it could be worth a try. Maybe.
    – user18033
    Commented Jan 14, 2018 at 14:00
  • @DonQuiKong I think you're right, insofar as the result is unexpected. That is probably the critical difference. If the number of variables were each chosen carefully to produce an unexpected result, that would be a good argument for allowability. In the absence of this, I would still say that the skilled person would select from among the available options to obtain the expected result in an obvious way. Or as European examiners harshly put it, they would be "mere implementation details".
    – Maca
    Commented Jan 14, 2018 at 19:30
  • one would probably have to take it to a high instance. And as having protection for exactly a specific set of values and having them published in return isn't worth it, we'll never know.
    – user18033
    Commented Jan 14, 2018 at 20:00

Depending on the wording of the claim as a whole it is possible but any claim with the specifics of weights and, correspondingly, a specific architecture of nodes would be easy to get around. As an example of language in a claim of an issued patent -

a trained generalized regression neural network trained using test reservoir data and test relative permeability values, with the trained generalized regression neural network for processing the actual reservoir data to determine a relative permeability prediction of an actual relative permeability in the hydrocarbon reservoir from the actual reservoir data; and

It is specific as to application and the source of the data but not to the gritty details of the neural network. One could imagine a dependent claim that nailed down the characteristics of an example network to greater and greater detail, including weights. It would not be the weights that gave patentablity.

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .