Context-adaptive subspace selection

Determine how to adapt global subspace selection for structured corrections to Gaussian-process posteriors when the leading directions of discrepancy vary strongly across conditioning contexts.

Background

The framework selects a single correction subspace in which to learn a mixture or diffusion law, while retaining Gaussian uncertainty in the complementary directions. This global subspace is effective in the reported experiments, but the authors note that different conditioning contexts may exhibit discrepancies in different directions. A method capable of adapting the selected subspace to context-dependent discrepancy structure would address an explicitly identified limitation of the framework.