BONUS TWEET Be sure to also check out this other paper posted at the same time, by @thesasho and Haohua Tang, also focused on unbiased algorithms in differential privacy. Despite similarities in the titles, the settings are mostly different. https://
arxiv.org/abs/2301.13850 9/8
MACHINE LEARNING
-
Unbiased Algorithms in Differential Privacy Research
By
–
-
Congratulations to Junior Researchers on Successful AI Paper
By
–
Anyway, this was a very satisfying paper (
https://
x.com/shortstein/sta
tus/1620279198333149184
…), and I think we really solved everything we set out to. Congrats to junior researchers @argymouz Matthew Regehr @vkerdos on this nice paper! https://
arxiv.org/abs/2301.13334 8/8 -
Differential Privacy Limitations for Gaussian Distribution Estimation
By
–
But we also show that pure (epsilon, 0)-DP is hopeless, even for really simple classes like Gaussians, and that the delta in approx DP is needed to perform unbiased estimation. 7/n
-
Unbiased Estimators for Symmetric Distributions
By
–
So that's sad. Is there any hope for special cases? If we happen to know the underlying distribution is symmetric, then *yes*, we can get unbiased estimators. 6/n
-
Unbiased Estimators: Better Algorithms for ML
By
–
This can be undesirable in a number of settings, when unbiased estimators are preferred. E.g., then we can compute the statistic multiple times on independent datasets and average them to reduce error. Natural question: are there better algorithms with no (or low) bias? 4/n
-
Balancing Bias and Variance in Statistical Estimation Methods
By
–
There's a problem though. By clipping the tails, this approach introduces significant (statistical) bias. Indeed, as we analyze in a paper with @vkerdos @thejonullman (
https://
arxiv.org/abs/2002.09464), we get minimax rates for estimation by *balancing* the bias & variance from noise. 3/n -
Private Mean Estimation: Clipping, Noise, and Differential Privacy
By
–
By now, we understand private mean estimation pretty well, especially in 1D. It turns out that one of the simplest algorithms, taking the empirical mean of the clipped samples (to restrict sensitivity) and adding noise (to introduce privacy) works pretty well. 2/n
-
Bias-Variance-Privacy Trilemma in Statistical Estimation
By
–
"A Bias-Variance-Privacy Trilemma for Statistical Estimation," with @argymouz
, Matthew Regehr, @vkerdos, @shortstein
, and @thejonullman
. https://
arxiv.org/abs/2301.13334 Private estimators MUST be biased! 1/n -
AlphaFold Reaches 1M+ Researchers in 18 Months
By
–
Big milestone: 1M+ researchers have now made use of #AlphaFold in their vital work, in the 18 months since we made it freely available.
— Demis Hassabis (@demishassabis) 1 février 2023
From antibiotic resistance to crop sustainability, it's been amazing to see the impact of ‘Science at Digital Speed’: https://t.co/f3SwqXBft8 pic.twitter.com/Tp6XJCJhPFBig milestone: 1M+ researchers have now made use of #AlphaFold in their vital work, in the 18 months since we made it freely available. From antibiotic resistance to crop sustainability, it's been amazing to see the impact of ‘Science at Digital Speed’: https://
unfolded.deepmind.com/#stories -
Getting Started with ML: Step-by-Step Guide and Open-Source Contribution
By
–
This is a perfect step-by-step guide for anyone getting started with ML. Along with this, I would also suggest joining open-source communities in the ML Space and start contributing to that to accelerate their learning.