Affordable Access

Spectral Experts for Estimating Mixtures of Linear Regressions

Authors
  • Chaganty, Arun Tejasvi
  • Liang, Percy
Type
Preprint
Publication Date
Jun 16, 2013
Submission Date
Jun 16, 2013
Identifiers
arXiv ID: 1306.3729
Source
arXiv
License
Yellow
External links

Abstract

Discriminative latent-variable models are typically learned using EM or gradient-based optimization, which suffer from local optima. In this paper, we develop a new computationally efficient and provably consistent estimator for a mixture of linear regressions, a simple instance of a discriminative latent-variable model. Our approach relies on a low-rank linear regression to recover a symmetric tensor, which can be factorized into the parameters using a tensor power method. We prove rates of convergence for our estimator and provide an empirical evaluation illustrating its strengths relative to local optimization (EM).

Report this publication

Statistics

Seen <100 times