Abstract
Learning using privileged information is an attractive problem setting thathelps many learning scenarios in the real world. A state-of-the-art method ofGaussian process classification (GPC) with privileged information is GPC+,which incorporates privileged information into a noise term of the likelihood.A drawback of GPC+ is that it requires numerical quadrature to calculate theposterior distribution of the latent function, which is extremelytime-consuming. To overcome this limitation, we propose a novel classificationmethod with privileged information based on Gaussian processes, called"soft-label-transferred Gaussian process (SLT-GP)." Our basic idea is that weconstruct another learning task of predicting soft labels (continuous values)obtained from privileged information and we perform transfer learning from thistask to the target task of predicting hard labels. We derive a PAC-Bayesianbound of our proposed method, which justifies optimizing hyperparameters by theempirical Bayes method. We also experimentally show the usefulness of ourproposed method compared with GPC and GPC+.