Estimating low-rank matrices via approximate message passing

Add to your list(s) Download to your calendar using vCal

• Ramji Venkataramanan, CUED
• Thursday 15 November 2018, 15:00-16:00
• LT6, Baker Building, CUED.

If you have a question about this talk, please contact Prof. Ramji Venkataramanan.

Large datasets often have an underlying low-dimensional structure that can be captured by modeling the data matrix as the sum of a low-rank matrix and a noise matrix. The goal is to estimate the low-rank part from the data matrix. A natural approach for estimating the low-rank part is via the spectrum of the data matrix. However, if the empirical distribution of the entries in the low-rank part is known, one can design estimators that substantially outperform simple spectral approaches.

In this talk we discuss an estimator that consists of an Approximate Message Passing (AMP) algorithm initialized with a spectral estimate. We obtain a sharp asymptotic characterization of the performance of this estimator, and use the result to derive detailed predictions for estimating a rank-one matrix and a block-constant low-rank matrix in Gaussian noise. Special cases of these models are closely related to the problem of community detection in stochastic block models. We show how the proposed estimator can be used to construct confidence intervals, and find that in many cases of interest, it can achieve Bayes-optimal accuracy above the spectral threshold.

(The talk will be self-contained, and will not assume familiarity with message passing algorithms.)

This talk is part of the Signal Processing and Communications Lab Seminars series.

This talk is included in these lists:

• All Talks (aka the CURE list)
• Cambridge Centre for Data-Driven Discovery (C2D3)
• Cambridge University Engineering Department Talks
• Cambridge talks
• Centre for Smart Infrastructure & Construction
• Chris Davis' list
• Computational Continuum Mechanics Group Seminars
• Featured lists
• Information Engineering Division seminar list
• Interested Talks
• LT6, Baker Building, CUED
• School of Technology
• Signal Processing and Communications Lab Seminars
• Trust & Technology Initiative - interesting events
• bld31
• ndk22's list
• ob366-ai4er
• rp587

Note that ex-directory lists are not shown.