Well, this is both fascinating and terrifying: https://arxiv.org/abs/2012.07805
This is honestly like turning the crank on a meat grinder backwards and having that work so well that that cows just mosey out of it under their own power. It is astounding.
@zeroed One reason for optimism is that it makes auditing for bias in training data a theoretical possibility, rather than leaving ML algorithms as a quasi-mystical blackboxes without verifiability or recourse.
Beautifully put.
Systematic bias can turn out devastating with such "scale" and the infamous "Security through obscurity" IMHO does not even get closer to be a point with this regard.
@mhoye the black box isn't that black after all!
@mhoye
Thank you for linking this paper.
I share the peculiar mix of feelings tho.
"It has become common to publish large (billion parameter) language models that have been trained on private datasets. This paper demonstrates that in such settings, an adversary can perform a training data extraction attack to recover individual training examples by querying the language model."
#cryptography and
#security (cs.CR); #computation and Language (cs.CL); #machinelearning (cs.LG)