A Discriminative Nonparametric Bayesian Model: Infinite Hidden Conditional Random Fields
Konstantinos Bousmalis, Louis‐Philippe Morency, Stefanos P. Zafeiriou, Maja Pantić · Neural Information Processing Systems · 2011
Finite Hidden Conditional Random Fields (HCRFs) [5] are discriminative models that learn the joint distribution of a class label and a sequence of latent variables conditioned on a given observation sequence, with dependencies among latent variables expressed by an undirected graph. A limitation of the finite HCRFs is that finding the optimal number of hidden states for a given classification problem is not always intuitive, and involves cross–validation, that can be very computationally expensive. This limitation motivated our nonparametric HCRF model that automatically learns the optimal number of hidden states given a specific dataset. This is achieved by using Hierarchical Dirichlet Processes (HDPs) to allow for an infinite number of hidden states for the HCRF. The reader is encouraged to look at [6] for a complete description of Hierarchical Dirichlet Processes [6].