10 ms·
To actually accomplish something like this purports to be (the linked tool only searches for similar images and doesn't tell you anything about how information
by whatshisface 4y ago
To actually accomplish something like this purports to be (the linked tool only searches for similar images and doesn't tell you anything about how information ended up inside the model), you could try removing individual images or sets of images from the same artist from the training dataset to see what outputs the resulting model would lose the ability to create. It would be expensive to do that for more than a few images, but given how helpful it would be for the debate about copyright and AI, I think it would be great if some researchers could try it.
- version_five 4y agoYour idea is the gold standard in explaining the influence of training data. People may be interested in this paper and more modern variations: https://arxiv.org/abs/1703.04730 https://arxiv.org/abs/1703.04730 It attempts to do as you suggest in a tractable way, to understand which training data is most influential.
- whatshisface 4y agoHas the output of this tool been measured against the gold standard so that we can tell whether or not it is working?
- version_five 4y agoNo, this tool (Stable Attribution) doesn't actually do training sample attribution. See my other comment https://news.ycombinator.com/item?id=34670483 https://news.ycombinator.com/item?id=34670483
- cypress66 4y agoAre these models actually "stable" (heh)? Or would changing anything in the dataset result in a butterfly effect so to speak, thus a completely different output for the same prompt?
- nickvincent 4y agoIt's an open challenge, this preprint proposes a solution for robust data valuation. https://arxiv.org/abs/2205.15466 https://arxiv.org/abs/2205.15466 However, I'm not aware of anyone actively trying to adapt these techniques for large generative AI models (though would be great to see).
- paconbork 4y agoSounds a bit like differential privacy