Abstract:
|
A conditional Gaussian (CG) distribution models a mixture of discrete and continuous variables in which, for each configuration of the discrete variables, the continuous variables have a multivariate Gaussian distribution. We present methodology for computing the globally most likely configuration of unobserved variables, given observed variables of a CG distribution. We represent the CG distribution, using ideas from graphical models, as a product of potentials over cliques in a junction tree. In this setting, we perform dynamic programming by representing the probability of the most likely configuration of each subtree in a form that is closed under the dynamic programming recursion. A "thinning" algorithm keeps the complexity of the representation computationally manageable, without involving approximation. We demonstrate an application to simultaneous estimation of time-varying tempo and rhythm, given musical performance data. In this application we estimate the MAP configuration of several thousand variables.
|