Figure 1

Illustration of Hierarchical Gaussian Process Regression. Observations from three hypothetical patients are plotted with a different marker. Lines are predicting mean functions, shaded areas are 78% credible intervals33 (predictive standard deviation) for the posterior. Three prediction errors are remarked. Left: a one-level model fits one distribution shared by all the samples, leading to high errors in individuals that are far from the mean. Right: in a Hierarchical GP, each patient follows an individual distribution (colors), and all these distributions follow an upper-level overall distribution (dotted lines), dramatically reducing the error.