An Acceleration Scheme for Memory Limited, Streaming PCA

  • 2018-07-17 16:25:01
  • Salaheddin Alakkari, John Dingliana
  • 2

Abstract

In this paper, we propose an acceleration scheme for online memory-limitedPCA methods. Our scheme converges to the first $k>1$ eigenvectors in a singledata pass. We provide empirical convergence results of our scheme based on thespiked covariance model. Our scheme does not require any predefined parameterssuch as the eigengap and hence is well facilitated for streaming datascenarios. Furthermore, we apply our scheme to challenging time-varying systemswhere online PCA methods fail to converge. Specifically, we discuss a family oftime-varying systems that are based on Molecular Dynamics simulations wherebatch PCA converges to the actual analytic solution of such systems.

 

Quick Read (beta)

loading the full paper ...