Latent variable model estimation apparatus, and method

摘要

To provide a latent variable model estimation apparatus capable of implementing the model selection at high speed even if the number of model candidates increases exponentially as the latent state number and the kind of the observation probability increase. A variational probability calculating unit 71 calculates a variational probability by maximizing a reference value that is defined as a lower bound of an approximation amount, in which Laplace approximation of a marginalized log likelihood function is performed with respect to an estimator for a complete variable. A model estimation unit 72 estimates an optimum latent variable model by estimating the kind and a parameter of the observation probability with respect to each latent state. A convergence determination unit 73 determines whether a reference value, which is used by the variational probability calculating unit 71 to calculate the variational probability, converges.

主权项

1. A latent variable model estimation apparatus comprising:
a variational probability calculating unit implemented at least by a hardware including a processor and that calculates a variational probability by maximizing a reference value, the reference value being defined as a lower bound of an approximation amount in which Laplace approximation of a marginalized log likelihood function is performed with respect to an estimator for a complete variable; a model estimation unit implemented at least by a hardware including a processor and that estimates an optimum latent variable model by estimating a kind and a parameter of an observation probability in each latent state; a convergence determination unit implemented at least by a hardware including a processor and that determines whether the reference value, which is used to calculate the variational probability by the variational probability calculating unit, converges; and an optimum model selection unit implemented at least by a hardware including a processor and that selects a latent variable model as the optimum latent variable model, the latent variable model corresponding to the larger reference value when the reference value converges in repeating loop processing in which the variational probability calculating unit calculates the variational probability, the model estimation unit estimates the optimum latent variable model, and the convergence determination unit determines whether the reference value converges, and when the reference value in the current loop processing is compared to the reference value in the preceding loop processing.