List Decodable Learning via Sum of Squares
Abstract
In the list-decodable learning setup, an overwhelming majority (say a -fraction) of the input data consists of outliers and the goal of an algorithm is to output a small list of hypotheses such that one of them agrees with inliers. We develop a framework for list-decodable learning via the Sum-of-Squares SDP hierarchy and demonstrate it on two basic statistical estimation problems {\it Linear regression:} Suppose we are given labelled examples containing a subset of {\it inliers} that are drawn i.i.d. from standard Gaussian distribution in , where the corresponding labels are well-approximated by a linear function . We devise an algorithm that outputs a list of linear functions such that there exists some that is close to . This yields the first algorithm for linear regression in a list-decodable setting. Our results hold for any distribution of examples whose concentration and anticoncentration can be certified by Sum-of-Squares proofs. {\it Mean Estimation:} Given data points containing a subset of {\it inliers} that are drawn i.i.d. from a Gaussian distribution in , we devise an algorithm that generates a list of means such that there exists close to . The recovery guarantees of the algorithm are analogous to the existing algorithms for the problem by Diakonikolas \etal and Kothari \etal. In an independent and concurrent work, Karmalkar \etal \cite{KlivansKS19} also obtain an algorithm for list-decodable linear regression using the Sum-of-Squares SDP hierarchy.
Cite
@article{arxiv.1905.04660,
title = {List Decodable Learning via Sum of Squares},
author = {Prasad Raghavendra and Morris Yau},
journal= {arXiv preprint arXiv:1905.04660},
year = {2019}
}