Abstract

Multi-Task Gaussian processes (MTGPs) have shown a significant progress bothin expressiveness and interpretation of the relatedness between differenttasks: from linear combinations of independent single-output Gaussian processes(GPs), through the direct modeling of the cross-covariances such as spectralmixture kernels with phase shift, to the design of multivariate covariancefunctions based on spectral mixture kernels which model delays among tasks inaddition to phase differences, and which provide a parametric interpretation ofthe relatedness across tasks. In this paper we further extend expressivenessand interpretability of MTGPs models and introduce a new family of kernelscapable to model nonlinear correlations between tasks as well as dependenciesbetween spectral mixtures, including time and phase delay. Specifically, we usegeneralized convolution spectral mixture kernels for modeling dependencies atspectral mixture level, and coupling coregionalization for discovering tasklevel correlations. The proposed kernels for MTGP are validated on artificialdata and compared with existing MTGPs methods on three real-world experiments.Results indicate the benefits of our more expressive representation withrespect to performance and interpretability.

 

Introduction (beta)

None

 

Conclusion (beta)

None