1: \begin{abstract}

2: 

3: Discriminative latent-variable models are typically learned using

4: EM or gradient-based optimization, which suffer from local optima.

5: In this paper, we develop a new computationally efficient

6: and provably consistent estimator for a mixture of linear regressions,

7: a simple instance of a discriminative latent-variable model.

8: Our approach relies on a low-rank linear regression to recover

9: a symmetric tensor, which can be factorized into the parameters

10: using a tensor power method.

11: We prove rates of convergence for our estimator

12: and provide an empirical evaluation illustrating

13: its strengths relative to local optimization (EM).

14: 

15: %  Spectral algorithms for latent variable models have seen considerable

16: %  recent interest for being efficient consistent estimators of model

17: %  parameters. These algorithms make few if any assumptions about the

18: %  generative process of the data, while providing a polynomial sample

19: %  and computational complexity. We present a new spectral algorithm for

20: %  a discriminative model, a mixture of linear regressions and show that

21: %  it can recover the regression coefficients with similar polynomial

22: %  guarantees. We evaluate the algorithm on linearly and non-linearly

23: %  generated data, as well as on a motion tracking task and compare it's

24: %  characteristics with an E-M algorithm.

25: \textbf{Last Modified: \today}

26: \end{abstract}