Figure 4
From: Deep Bayesian Gaussian processes for uncertainty estimation in electronic health records
Predictive coverage for DKL and DBGP. The red path and the blue path represent DKL and DBGP, respectively. DKL maps the raw input x to the latent features, which is 9.5 in the figure, using a deterministic model \(f_{d}\). Thus, the GP regressor predicts y for the given \(f_{d}(x)=9.5\), and the potential prediction y is represented as the yellow line. In contrast, DBGP maps the input x to the latent distribution using a probabilistic model \(f_{p}\), and the GP regressor predicts y conditioned on the latent distribution \(f_{p}(x)\). The potential prediction y is represented by the yellow area. Typically, DKL and DBGP make prediction by marginalising the yellow line and the yellow area, respectively. Only one GP regressor is shown here to simplify the description. In practice, both GP regressor and \(f_{p}\) or \(f_{d}\) are trained together.
