Learning Gaussian processes from multiple tasks

We consider the problem of multi-task learning, that is, learning multiple related functions. Our approach is based on a hierarchical Bayesian framework, that exploits the equivalence between parametric linear models and nonparametric Gaussian processes (GPs). The resulting models can be learned easily via an EM-algorithm. Empirical studies on multi-label text categorization suggest that the presented models allow accurate solutions of these multi-task problems.

• Type: Conference
• Year: 2005
• Where: ICML
• Authors: Kai Yu, Volker Tresp, Anton Schwaighofer
• Keywords: ICML 2005, Machine Learning, Multi-task Learning, Multi-task Problems, Parametric Linear Models