What are the implications of patents on deep learning techniques? What are the limitations when using such techniques? For example, dropout or transformers are patented by Google (https://patents.google.com/patent/US9406017B2/en, https://patents.google.com/patent/US10452978B2/en), yet they are implemented in big libraries like PyTorch (https://pytorch.org/docs/stable/generated/torch.nn.Dropout.html, https://pytorch.org/docs/stable/generated/torch.nn.Transformer.html) and used in many projects.
4
-
1I'm afraid such open ended questions aren't well suited to this site. In any case it isn't clear the linked pytorch libraries are infringing on the cited google patents. For instance the Transformer feature cites a paper which predates the relevant Google patent by at least a year. Dropout seems more problematic. I found the following links that seems relevant. medium.com/syncedreview/… analyticsindiamag.com/…– Eric S ♦Jan 5 at 18:07
-
Thank you for the links! How is the time difference you mention relevant? The inventors of the patent are the same as the authors of the paper. Do you know a site better suited to ask this question? I searched the internet and this site is the best one I could find.– salccJan 5 at 18:18
-
Public disclosure usually disqualifies you from pursuing a patent. In the US there is a one year grace period, but large companies rarely disclose a patentable technology before filing a patent application.– Eric S ♦Jan 5 at 23:58
-
I don't know of a better site to pose your question. Here at Ask Patents, there are some experts in the process of obtaining patents, but no one specifically expert in machine learning. A better question is whether Google plans on enforcing the patents cited. This is a business decision and beyond the scope of this site too.– Eric S ♦Jan 6 at 1:25
Add a comment
|