NP-PROV: Neural Processes with Position-Relevant-Only Variances

Neural Processes (NPs) families encode distributions over functions to a latent representation, given context data, and decode posterior mean variance at unknown locations. Since are derived from the same space, they may fail on out-of-domain tasks where fluctuations in function values amplify model uncertainty. We present new member named with Position-Relevant-Only Variances (NP-PROV). NP-PROV hypothesizes that target point close has small uncertainty, regardless of value position. The resulting approach derives function-value-related space position-related-only separately. Our evaluation synthetic real-world datasets reveals can achieve state-of-the-art likelihood while retaining bounded when drifts exist value.