An Inertial Latent-Variable Sequence Model.
Rajasekaran Masatran · arXiv (Cornell University) · 2015
Latent-variable models are one popular approach to modeling sequences. One problem with sequence models, including latent-variable models, is that their exact learning algorithms are usually intractable in T, the length of the sequence, necessitating the use of approximation algorithms. Though these algorithms are faster than their exact counterparts, they are iterative and computationally expensive. However, models with fast algorithms can be designed for commonly occurring subsets of the set of sequences. We propose a new statistical model for a subset---the set of sequences with inertia. Our learning algorithms, at time complexity O(T log T), are significantly faster than those of general-purpose latent-variable sequence models.