Gaussian Process Model Cuts Overconfidence by 20%

Researchers Yean Hoon Ong, Paolo Barucca, Wei Pan and Jun Wang announced on August 15 2026 that they have developed a novel variant of the product‑of‑experts Gaussian process (GP‑pro) called GP‑pro‑c, which incorporates an information‑based method to recalibrate overestimated posterior variances in local experts. Their work appears in the Journal of Artificial Intelligence Research (JAIR).¹

The motivation stems from a long‑standing issue in Gaussian processes: while a single global GP offers accurate predictions, its cubic computational cost limits scalability to large datasets. Product‑of‑experts models mitigate this by training multiple local GPs on disjoint data subsets, then combining their outputs; however, because each expert is trained independently, the resulting combined posterior often shows inflated confidence—posterior variances are systematically too low.¹

To address this, the team introduced a calibration ratio derived from the monotonicity and submodularity properties of information gain within Gaussian processes. By measuring how much additional data each local expert contributes, GP‑pro‑c scales down the variance estimate for each expert proportionally, thereby reducing false confidence estimates.¹

The performance of GP‑pro‑c was evaluated using three metrics: negative log‑likelihood (NLL), root mean squared error (RMSE), and expected normalised ca—a measure capturing overall predictive uncertainty.¹

Results show a roughly 20 percent reduction in false confidence compared with conventional product‑of‑expert GPs, suggesting significant practical gains where accurate uncertainty is critical.

The study also notes that further validation against diverse, large‑scale datasets will determine whether the observed improvement persists beyond controlled environments.

In sum, Ong, Barucca, Pan and Wang have provided a concrete solution to a persistent scalability–uncertainty trade‑off in Gaussian processes, offering both theoretical insight and empirical evidence of a ~20 percent reduction in overconfidence when applying their GP‑pro‑c framework. The work demonstrates how principled use of information‑theoretic concepts can refine machine‑learning models without sacrificing computational efficiency.¹