Ask HN: Are weights trained from a dataset also under the same license terms?
A lot of the datasets that are commonly used nowadays are bound by some license or the other. ImageNet for example - http://image-net.org/download-faq
I understand when the use of the actual images is restricted under the terms, but what about weights that we derive out of millions of these images? There is no actual trace of the image within the trained networks, its just a model derived by various iterations over the images. Much like we humans read books and derive knowledge (Most books by themselves are bound by restrictive copyright, but we are free to apply the knowledge we derive from them).
How do companies deal with this? Most companies using deep learning will have to use large datasets and actually creating the data(for example, of millions of images for tens/hundreds of classes) would not be viable.
Also when we do transfer learning on these heavily trained models, we essentially change the network, what about the licenses then?