A recursive Bayesian approach to multitask learning
by
Francesco Dinuzzo
Università di Pavia
Coauthors: Gianluigi Pillonetto, Giuseppe De Nicolao

Multi-task learning deals with the simultaneous estimation of several related functions (tasks) from empirical data. Recent studies have shown that exploiting similarities between the tasks it is possible to improve the generalization performance with respect to the standard single-task approach which learns the tasks separately. In this paper, we address two major open issues, namely the reduction of computational complexity and the development of an efficient recursive algorithm to be used for online applications were data are sequentially processed. The problem is formulated and solved within a Bayesian setting modelling the unknown tasks as Gaussian random fields. The potential advantage of the multi-task approach with respect to the single task one and the reduction of computational complexity are assessed on simulated data as well as on a real pharmacokinetic experiment.

Date received: March 28, 2008