Self-Averaging Expectation Propagation

Burak Cakmak, Manfred Opper, Bernard Henri Fleury, Ole Winther

Research output: Contribution to conference without publisher/journalPosterResearchpeer-review

61 Downloads (Pure)

Abstract

We investigate the problem of approximate inference using Expectation Propagation
(EP) for large systems under some statistical assumptions. Our approach tries
to overcome the numerical bottleneck of EP caused by the inversion of large
matrices. Assuming that the measurement matrices are realizations of specific
types of random matrix ensembles – called invariant ensembles – the EP cavity
variances have an asymptotic self-averaging property. They can be pre-computed
using specific generating functions which do not require matrix inversions. We
demonstrate the performance of our approach on a signal recovery problem of
compressed sensing and compare with standard EP.
Original languageEnglish
Publication date2016
Number of pages5
Publication statusPublished - 2016
EventAdvances in Approximate Bayesian Inference : NIPS 2016 Workshop -
Duration: 9 Dec 20169 Dec 2016
http://approximateinference.org/

Workshop

WorkshopAdvances in Approximate Bayesian Inference
Period09/12/201609/12/2016
Internet address

Fingerprint

Recovery

Cite this

Cakmak, B., Opper, M., Fleury, B. H., & Winther, O. (2016). Self-Averaging Expectation Propagation. Poster presented at Advances in Approximate Bayesian Inference , .
Cakmak, Burak ; Opper, Manfred ; Fleury, Bernard Henri ; Winther, Ole. / Self-Averaging Expectation Propagation. Poster presented at Advances in Approximate Bayesian Inference , .5 p.
@conference{3ab6fb53e3c8448fa86416b366ae6872,
title = "Self-Averaging Expectation Propagation",
abstract = "We investigate the problem of approximate inference using Expectation Propagation(EP) for large systems under some statistical assumptions. Our approach triesto overcome the numerical bottleneck of EP caused by the inversion of largematrices. Assuming that the measurement matrices are realizations of specifictypes of random matrix ensembles – called invariant ensembles – the EP cavityvariances have an asymptotic self-averaging property. They can be pre-computedusing specific generating functions which do not require matrix inversions. Wedemonstrate the performance of our approach on a signal recovery problem ofcompressed sensing and compare with standard EP.",
author = "Burak Cakmak and Manfred Opper and Fleury, {Bernard Henri} and Ole Winther",
year = "2016",
language = "English",
note = "Advances in Approximate Bayesian Inference : NIPS 2016 Workshop ; Conference date: 09-12-2016 Through 09-12-2016",
url = "http://approximateinference.org/",

}

Cakmak, B, Opper, M, Fleury, BH & Winther, O 2016, 'Self-Averaging Expectation Propagation', Advances in Approximate Bayesian Inference , 09/12/2016 - 09/12/2016.

Self-Averaging Expectation Propagation. / Cakmak, Burak; Opper, Manfred ; Fleury, Bernard Henri; Winther, Ole.

2016. Poster presented at Advances in Approximate Bayesian Inference , .

Research output: Contribution to conference without publisher/journalPosterResearchpeer-review

TY - CONF

T1 - Self-Averaging Expectation Propagation

AU - Cakmak, Burak

AU - Opper, Manfred

AU - Fleury, Bernard Henri

AU - Winther, Ole

PY - 2016

Y1 - 2016

N2 - We investigate the problem of approximate inference using Expectation Propagation(EP) for large systems under some statistical assumptions. Our approach triesto overcome the numerical bottleneck of EP caused by the inversion of largematrices. Assuming that the measurement matrices are realizations of specifictypes of random matrix ensembles – called invariant ensembles – the EP cavityvariances have an asymptotic self-averaging property. They can be pre-computedusing specific generating functions which do not require matrix inversions. Wedemonstrate the performance of our approach on a signal recovery problem ofcompressed sensing and compare with standard EP.

AB - We investigate the problem of approximate inference using Expectation Propagation(EP) for large systems under some statistical assumptions. Our approach triesto overcome the numerical bottleneck of EP caused by the inversion of largematrices. Assuming that the measurement matrices are realizations of specifictypes of random matrix ensembles – called invariant ensembles – the EP cavityvariances have an asymptotic self-averaging property. They can be pre-computedusing specific generating functions which do not require matrix inversions. Wedemonstrate the performance of our approach on a signal recovery problem ofcompressed sensing and compare with standard EP.

UR - http://approximateinference.org/accepted/

M3 - Poster

ER -

Cakmak B, Opper M, Fleury BH, Winther O. Self-Averaging Expectation Propagation. 2016. Poster presented at Advances in Approximate Bayesian Inference , .