Compressive Independent Component Analysis.

Mikey Sheehan, Madeleine Kotzagiannidis, Michael Davies

Research output: Contribution to conferencePaperpeer-review

Abstract

In this paper we investigate the minimal dimensionstatistic necessary in order to solve the independent componentanalysis (ICA) problem. We create a compressive learning frame-work for ICA and show for the first time that the memorycomplexity scales only quadratically with respect to the numberof independent sourcesn, resulting in a vast improvement overother ICA methods. This is made possible by demonstrating alow dimensional model set, that exists in the cumulant basedICA problem, can be stably embedded into a compressed spacefrom a larger dimensional cumulant tensor space. We show thatidentifying independent source signals can be achieved with highprobability when the compression sizemis of the optimal orderof the intrinsic dimension of the ICA parameters and propose aiterative projection gradient algorithm to achieve this.
Original languageEnglish
Number of pages5
Publication statusPublished - 2 Sep 2019
Event27th European Signal Processing Conference - A Coruna, Spain
Duration: 2 Sep 20196 Sep 2019

Conference

Conference27th European Signal Processing Conference
Abbreviated titleEUSIPCO 2019
CountrySpain
CityA Coruna
Period2/09/196/09/19

Fingerprint Dive into the research topics of 'Compressive Independent Component Analysis.'. Together they form a unique fingerprint.

Cite this