Orateur
Description
Multitask gaussian processes are popular tools for learning several correlated outputs, and find applications for instance in medicine, robotics, earth sciences, etc. In this field, the Linear Model of Co-regionalization (LMC) is a very general model, which expressivity and conceptual simplicity are appealing; however, its cubic complexity in both the number of datapoints and number of tasks makes exact computation impractical for most applications, making simplifications or approximations - in general quite complex - mandatory.
We here show that under a very mild restriction on the structure of the noise model, the LMC can actually decouple over latent processes, leading to a complexity that is only linear in the number of said processes. We show how to parametrize and optimize the resulting model, and confirm its excellent behavior with a parametric study on synthetic data. We finally apply this work to a problem of neutronics simulation : the reconstruction of homogenized cross-sections in deterministic codes.
