v4dok

17 ms·

151.
by v4dok 7y ago
Care to share more info?
152.
by v4dok 7y ago
I believe that the only way to effectively protect privacy in a world where the effects of the aggregate privacy violation are severe, but individually privacy is not as "important" (i.e most of the people don't care enough t
153.
by v4dok 7y ago
MPC works only hand-in-hand with DP. MPC ensures that you don't have access to the "remote" dataset, but does nothing to mitigate model "memorising" specific private information (if we are talking about advanced ana
154.
by v4dok 7y ago
Not at all, the very act of learning from that data was (arguably) illegal. Reading a book is not illegal.
155.
by v4dok 7y ago
But if I bought a book, read it, and then recited from memory as good as I could (which is obviously not 1:1, but a vague idea of the story), adding things from other books where I didn't really remember what was happening. I think, I&
156.
by v4dok 7y ago
If you don't have 1:1 match in the text, I see no reason why it would be considered an infrigement. Isn't it what people also do? There is no parthenogenesis so a system that "learns" and doesn't reproduce should be
157.
by v4dok 7y ago
Maybe I don't understand correctly but my point is that BERT and friends just let you skip a lot of feature engineering. Which(the F.Eng) is actually pretty useless once you shift the corpus a bit. That is for me the biggest contributi
158.
by v4dok 7y ago
What I have seen is that one of the main issues in NLP is actually finding annotated text. Especially for some more specialized tasks annotating can be a very costly process. It is not easy to annotate an e.g an argument versus labelling a
159.
by v4dok 7y ago
Can you give a bit more info on that? What is not accurate?