Provable ICA with Unknown Gaussian Noise, and Implications for Gaussian Mixtures and Autoencoders
Author(s): Arora, Sanjeev; Ge, Rong; Moitra, Ankur; Sachdeva, Sushant
DownloadTo refer to this page use:
http://arks.princeton.edu/ark:/88435/pr1885k
Abstract: | We present a new algorithm for independent component analysis which has provable performance guarantees. In particular, suppose we are given samples of the form π¦=π΄π₯+π where π΄ is an unknown but non-singular πΓπ matrix, π₯ is a random variable whose coordinates are independent and have a fourth order moment strictly less than that of a standard Gaussian random variable and π is an π-dimensional Gaussian random variable with unknown covariance π΄: We give an algorithm that provably recovers π΄ and π΄ up to an additive π and whose running time and sample complexity are polynomial in π and 1/π. To accomplish this, we introduce a novel βquasi-whiteningβ step that may be useful in other applications where there is additive Gaussian noise whose covariance is unknown. We also give a general framework for finding all local optima of a function (given an oracle for approximately finding just one) and this is a crucial step in our algorithm, one that has been overlooked in previous attempts, and allows us to control the accumulation of error when we find the columns of π΄ one by one via local search. |
Publication Date: | 2015 |
Citation: | Arora, Sanjeev, Rong Ge, Ankur Moitra, and Sushant Sachdeva. "Provable ICA with Unknown Gaussian Noise, and Implications for Gaussian Mixtures and Autoencoders." Algorithmica 72, no. 1 (2015): 215-236. doi:10.1007/s00453-015-9972-2 |
DOI: | 10.1007/s00453-015-9972-2 |
ISSN: | 0178-4617 |
EISSN: | 1432-0541 |
Pages: | 215 - 236 |
Type of Material: | Journal Article |
Journal/Proceeding Title: | Algorithmica |
Version: | Author's manuscript |
Items in OAR@Princeton are protected by copyright, with all rights reserved, unless otherwise indicated.