Neural diffusion processes are a scalable, non-Gaussian way to model distributions over functions, but existing versions only handle single-task inference and cannot capture dependencies between related tasks. The paper introduces multi-task neural diffusion processes (MT-NDPs), which extend the framework with a task encoder that extracts a low-dimensional representation from context observations and conditions the diffusion model on it. This allows information sharing across tasks while keeping the model flexible enough to handle context sets of varying size and invariant to the order of those observations.
The authors evaluate MT-NDPs on synthetic experiments and two real-world applications: wind turbine data and school examination outcomes. Compared with single-task diffusion baselines, MT-NDPs achieve improved predictive accuracy when context observations are available, and they perform competitively against neural-process and Gaussian-process alternatives. In the wind-energy application, MT-NDPs record the lowest average coverage error among all evaluated models at larger context sizes, indicating better-calibrated predictive uncertainty.
The results suggest that task-conditioned diffusion models are a promising direction for uncertainty-aware multi-task regression and few-shot adaptation to previously unseen tasks. The paper is a single source, so no conflicting findings are reported; the claims are based on the authors' own experiments and comparisons.