NeurIPS 2018, 32nd Neural Information Processing Systems Conference, 2-8 December, 2018, Montreal, Canada
Multi-fidelity models are prominently used in various science and engineering
applications where cheaply-obtained, but possibly biased and noisy observations
must be effectively combined with limited or expensive true data in order to construct
reliable models. The notion of applying deep Gaussian processes (DGPs)
to this setting has recently shown great promise by capturing complex nonlinear
correlations across fidelities. However, the architectures explored thus far are burdened
by structural assumptions and constraints which deter such models from performing
to the best of their expected capabilities. In this paper we propose a novel
approach for DGP multi-fidelity modeling which treats DGP layers as fidelity levels
and uses a variational inference scheme to propagate uncertainty across them.
In our experiments, we show that this approach makes substantial improvements
in quantifying and propagating uncertainty in multi-fidelity set-ups, which in turn
improves their effectiveness in decision-making pipelines.