Onsager-Machlup Posterior Transport for Deep Gaussian Processes
Researchers have introduced OM-Path, a novel method for approximate inference in Deep Gaussian Processes (DGPs). This approach frames inference as posterior transport, learning a deterministic sampler to map a reference measure to inducing variables. OM-Path utilizes the Onsager-Machlup action as a path regularizer and has shown statistically significant improvements over existing methods like DBVI on specific large datasets, though it performs comparably on smaller, noisier datasets. AI
IMPACT Introduces a novel inference technique that may enhance the performance of Deep Gaussian Processes on complex regression tasks.