4 ms·
Genuine question: why would they be dangerous? By 'model weights' I assume you mean the weights on a neural net, yes? (I skimmed the article and didn't see, er
by MikeTheGreat 3y ago
Genuine question: why would they be dangerous?
By 'model weights' I assume you mean the weights on a neural net, yes? (I skimmed the article and didn't see, er, any technical details, but I could have missed something :) ).
I thought that at this point pretty much nobody knows what neural net weights mean - it's kind of a 'math soup' that results from back-propagating adjustments to your huge linear equation based on training data.
That said - you're not the only person here saying this / agreeing with this.
What am I missing? :)
- CGamesPlay 3y agoYou use more math soup that gets the original examples out. Not perfectly, but possibly good enough. Reconstructing Training Data from Trained Neural Networks https://arxiv.org/abs/2206.07758 https://arxiv.org/abs/2206.07758
- MikeTheGreat 3y agoHuh. I never would have expected that. Every day I learn something new, I guess :) Thanks for the link!