26.10.2013 Views

Nonparametric Bayesian Discrete Latent Variable Models for ...

Nonparametric Bayesian Discrete Latent Variable Models for ...

Nonparametric Bayesian Discrete Latent Variable Models for ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

3.4 Dirichlet Process Mixtures of Factor Analyzers<br />

FA is generally used as a dimensionality reduction technique by assuming that the<br />

dimension of the latent factors z is much less than the data dimension, Q ≪ D. For<br />

this case, a FA can be interpreted as a Gaussian with a constrained covariance matrix.<br />

Assuming that most of the structure lies in a lower dimensional space, MFA can be used<br />

to model high dimensional data as a mixture of Gaussians with less computational cost.<br />

Analytical inference is not possible <strong>for</strong> the MFA models but maximum likelihood<br />

methods such as expectation maximization (EM) can be used to make point estimates<br />

<strong>for</strong> the parameters (Ghahramani and Hinton, 1996) or MCMC methods can be used<br />

<strong>for</strong> obtaining the posterior distribution (Utsugi and Kumagai, 2001). These algorithms<br />

are straight<strong>for</strong>ward to apply <strong>for</strong> fixed latent dimension and a fixed number of components.<br />

Deciding on the latent dimension and the number of components to be used<br />

is an important modeling decision <strong>for</strong> MFA which is usually dealt with by using cross<br />

validation. Alternatively, learning the latent dimension or the number of components<br />

can be included in the inference.<br />

<strong>Bayesian</strong> inference on the latent dimension of the FA model has been studied by Lopes<br />

and West (2004) using reversible jump MCMC (RJMCMC).<br />

Ghahramani and Beal (2000) <strong>for</strong>mulate a variational method <strong>for</strong> learning the MFA<br />

model. They <strong>for</strong>m a hierarchical MFA model by placing priors on the component parameters<br />

and hyperpriors on some of the hyperparameters and do inference on the model<br />

using a variational approximation to the log marginal likelihood. The dimension of<br />

hidden factors <strong>for</strong> each component is determined by automatic relevance determination<br />

(ARD) (MacKay, 1994; Neal, 1996), and the number of components are determined by<br />

birth-death moves. The drawback of this method is that it is very sensitive to random<br />

initialization of the parameters and since it is an approximation, it does not guarantee<br />

finding the exact solution.<br />

Fokoué and Titterington (2003) present a method <strong>for</strong> inferring both the latent dimension<br />

and the number of components in an MFA model using birth and death MCMC<br />

(BDMCMC) of Stephens (2000) which is an alternative to RJMCMC. They treat both<br />

the hidden dimension and the number of components as a parameter of the model and<br />

do inference on both using BDMCMC.<br />

For these approaches, the number of components can be seen as a parameter of the<br />

model which is inferred from the data. The DP prior allows a nonparametric <strong>Bayesian</strong><br />

<strong>for</strong>mulation of the MFA model, eliminating the need to do inference on the number of<br />

components necessary to represent the data.<br />

In this section, we introduce the Dirichlet process mixtures of factor analyzers model<br />

(DPMFA), a FA model with a DP prior over the distribution of the model parameters.<br />

Dirichlet Process Mixtures of Factor Analysers<br />

We start with the distributional assumptions of (Ghahramani and Beal, 2000) <strong>for</strong> the<br />

parametric MFA model, and <strong>for</strong>m the Dirichlet process MFA model by taking K → ∞.<br />

In detail, the prior <strong>for</strong> the means µ j is Gaussian,<br />

µ j ∼ N (ξ, R −1 ), (3.69)<br />

53

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!