Multi-Task Learning for Sequential Data

The problem of multi-task learning (MTL) is considered for sequential data, such as that typically modeled via a hidden Markov model (HMM). A given task is composed of a set of sequential data, for which an HMM is to be learned, and MTL is employed to learn the multiple task-dependent HMMs jointly, through appropriate sharing of data. The HMM-MTL formulation is implemented in a Bayesian setting, by utilizing a common prior on the cross-task HMM parameters. The prior is characterized in a nonparametric manner, utilizing a Dirichlet process (DP), and a variational Bayes (VB) formulation is employed for efficient inference. The DP-based HMM-MTL formulation is demonstrated using both synthesized and real sequential data, wherein the MTL formulation is demonstrated to yield improved performance relative to single-task learning, for cases in which at least a subset of the tasks are related, with this task relatedness determined automatically by the algorithm. I