We study the fundamental problem of learning the parameters of a
high-dimensional Gaussian in the presence of noise -- where an
$\varepsilon$-fraction of our samples were chosen by an adversary. We give
robust estimators that achieve estimation error $O(\varepsilon)$ in the total
variation distance, which is optimal up to a universal constant that is
independent of the dimension.
In the case where just the mean is unknown, our robustness guarantee is
optimal up to a factor of $2$ and the running time is polynomial in $d$
and $1/\epsilon$. When both the mean and covariance are unknown, the running
time is polynomial in $d$ and quasipolynomial in $1/\varepsilon$. Moreover all
of our algorithms require only a polynomial number of samples. Our work shows
that the same sorts of error guarantees that were established over fifty years
ago in the one-dimensional setting can also be achieved by efficient algorithms
in high-dimensional settings.
Description
[1704.03866] Robustly Learning a Gaussian: Getting Optimal Error, Efficiently
%0 Journal Article
%1 diakonikolas2017robustly
%A Diakonikolas, Ilias
%A Kamath, Gautam
%A Kane, Daniel M.
%A Li, Jerry
%A Moitra, Ankur
%A Stewart, Alistair
%D 2017
%K learning robustness stats
%T Robustly Learning a Gaussian: Getting Optimal Error, Efficiently
%U http://arxiv.org/abs/1704.03866
%X We study the fundamental problem of learning the parameters of a
high-dimensional Gaussian in the presence of noise -- where an
$\varepsilon$-fraction of our samples were chosen by an adversary. We give
robust estimators that achieve estimation error $O(\varepsilon)$ in the total
variation distance, which is optimal up to a universal constant that is
independent of the dimension.
In the case where just the mean is unknown, our robustness guarantee is
optimal up to a factor of $2$ and the running time is polynomial in $d$
and $1/\epsilon$. When both the mean and covariance are unknown, the running
time is polynomial in $d$ and quasipolynomial in $1/\varepsilon$. Moreover all
of our algorithms require only a polynomial number of samples. Our work shows
that the same sorts of error guarantees that were established over fifty years
ago in the one-dimensional setting can also be achieved by efficient algorithms
in high-dimensional settings.
@article{diakonikolas2017robustly,
abstract = {We study the fundamental problem of learning the parameters of a
high-dimensional Gaussian in the presence of noise -- where an
$\varepsilon$-fraction of our samples were chosen by an adversary. We give
robust estimators that achieve estimation error $O(\varepsilon)$ in the total
variation distance, which is optimal up to a universal constant that is
independent of the dimension.
In the case where just the mean is unknown, our robustness guarantee is
optimal up to a factor of $\sqrt{2}$ and the running time is polynomial in $d$
and $1/\epsilon$. When both the mean and covariance are unknown, the running
time is polynomial in $d$ and quasipolynomial in $1/\varepsilon$. Moreover all
of our algorithms require only a polynomial number of samples. Our work shows
that the same sorts of error guarantees that were established over fifty years
ago in the one-dimensional setting can also be achieved by efficient algorithms
in high-dimensional settings.},
added-at = {2020-02-26T13:41:35.000+0100},
author = {Diakonikolas, Ilias and Kamath, Gautam and Kane, Daniel M. and Li, Jerry and Moitra, Ankur and Stewart, Alistair},
biburl = {https://www.bibsonomy.org/bibtex/23b91a638fa7a0b90b14ed461bda28c57/kirk86},
description = {[1704.03866] Robustly Learning a Gaussian: Getting Optimal Error, Efficiently},
interhash = {db771bfbe907a2a34642c62614320e6c},
intrahash = {3b91a638fa7a0b90b14ed461bda28c57},
keywords = {learning robustness stats},
note = {cite arxiv:1704.03866Comment: To appear in SODA 2018},
timestamp = {2020-02-26T13:41:35.000+0100},
title = {Robustly Learning a Gaussian: Getting Optimal Error, Efficiently},
url = {http://arxiv.org/abs/1704.03866},
year = 2017
}