Multi-task learning (MTL) based on extensions of the Dirichlet process (DP) is considered for clustering and classification. This dissertation presents several generic non-parametric Bayesian models for inference in various data sets. In particular, we investigate the situation where multiple correlated data are collected using hierarchical structures, with the objective of learning multiple tasks in parallel in order to share the common information contained among tasks. We develop a new inference algorithm combining Markov chain Monte Carlo (MCMC) and variational Bayesian (VB) inference methods for a new model, called the hierarchical kernel stick-breaking process (H-KSBP), and demonstrate the robustness and speed of the algorithm on a benchmark image segmentation data set. Two other novel MTL learning models with distinct sharing mechanisms are proposed to perform classification in landmine detection and art image retrieval. A comprehensive analysis of performance is given for three data sets, with comparisons also made to other approaches, such as Xue et al. [XLCK07] MTL and single-task learning.
展开▼
