5 ms·

I think the discussions miss a bit an important point. IANAL, but I think if a young programmer reads a lot of source code on GitHub, and based on this reading

by _ph_ 5y ago

I think the discussions miss a bit an important point. IANAL, but I think if a young programmer reads a lot of source code on GitHub, and based on this reading becomes a better programmer, this is a fair use of copyrighted material and pretty much independant of the license. If I read any book and learn the corresponding language, this isn't a copyright violation of the book either. This starts, when I begin to quote from that book or the programmer takes snippets from the programs that got read. The problem is, I don't think you can really claim that Copilot learned to program. While some of the output seems to be something new, most of the times it looks more like a recomposition of learned fragments if not even longer pieces of verbatim code taken from copyrighted material. We have seen examples of this. And in this moment, it becomes a copyright discussion, probably determined by the volume of copyrighted material reproduced. Which by the way is always the risk if a human uses certain training material. The better one is at memorizing things, the more there is the risk. Or put it the other way around: if Copilot would use its "knowledge" of programs to advise the programmer like pointing out potential errors without reproducing anything it used for learning, it should be fine. But that is not how it works.

jozvolskyef 5y ago
If a person who's never seen a goat looks at a million copyrighted images of goats and draws a goat, are they committing copyright infringement? What if an algorithm does the same? The result is 'a recomposition of learned fragments' in either case.
- _ph_ 5y ago
  The point is, with a human you can assume that the person did learn how a goat looks and that a drawing isn't a mechanical translation of one image. Only in cases, where a single image is faithfully reproduced in the drawing, copyright questions might arise. But the base for this is: humans can learn and have an obvious creative process. With the AI it is not clear, that there is an abstract lerning beyond building certain memories.
  - jozvolskyef 5y ago
    From what I understand about these models, having studied and applied machine learning but not deep learning, they all build abstractions (embeddings) akin to eigenfaces[1][2] where many samples of a feature are turned into an abstract feature. These abstract features are then manipulated by the model. Sure, just like with Fourier transform-based compression, anything can be reconstructed from primitive features, it is called overfitting and the challenge of training these models is building a learning process that achieves generalization rather than overfitting. It is not true that 'it is not clear that there is abstract learning', because without it, generalization is impossible, and without generalization the model wouldn't pass even the most trivial benchmarks. [1]: https://media.geeksforgeeks.org/wp-content/uploads/20200317134836/train_faces.png https://media.geeksforgeeks.org/wp-content/uploads/202003171... [2]: https://www.geeksforgeeks.org/ml-face-recognition-using-eigenfaces-pca-algorithm/ https://www.geeksforgeeks.org/ml-face-recognition-using-eige...

by _ph_ 5y ago

jozvolskyef 5y ago

If a person who's never seen a goat looks at a million copyrighted images of goats and draws a goat, are they committing copyright infringement? What if an algorithm does the same? The result is 'a recomposition of learned fragments' in either case.

_ph_ 5y ago
The point is, with a human you can assume that the person did learn how a goat looks and that a drawing isn't a mechanical translation of one image. Only in cases, where a single image is faithfully reproduced in the drawing, copyright questions might arise. But the base for this is: humans can learn and have an obvious creative process. With the AI it is not clear, that there is an abstract lerning beyond building certain memories.
- jozvolskyef 5y ago
  From what I understand about these models, having studied and applied machine learning but not deep learning, they all build abstractions (embeddings) akin to eigenfaces[1][2] where many samples of a feature are turned into an abstract feature. These abstract features are then manipulated by the model. Sure, just like with Fourier transform-based compression, anything can be reconstructed from primitive features, it is called overfitting and the challenge of training these models is building a learning process that achieves generalization rather than overfitting. It is not true that 'it is not clear that there is abstract learning', because without it, generalization is impossible, and without generalization the model wouldn't pass even the most trivial benchmarks. [1]: https://media.geeksforgeeks.org/wp-content/uploads/20200317134836/train_faces.png https://media.geeksforgeeks.org/wp-content/uploads/202003171... [2]: https://www.geeksforgeeks.org/ml-face-recognition-using-eigenfaces-pca-algorithm/ https://www.geeksforgeeks.org/ml-face-recognition-using-eige...