Lunch colloquium Frida Viset (PhD at Data Driven Control)

05 October 2022 12:30 till 13:30 - Location: instruction room j, 3me - By: DCSC | Add to my calendar

"Fast Gaussian Process Predictions on Large Domains with Prediction-Point Dependent Basis Functions"

Concepts from sensor fusion can help remedy the computational challenges of performing online approximate Gaussian process predictions for large spatial fields. Specifically, concepts from Kalman filtering on the standard and the information form generalizes to spatially scalable parametric approximations to Gaussian process regression. 

Based on this interpretation, we can show that a subset of the trained system on information form is equivalent to a trained subset of the system. This notion can be used to give further computational gains in the Gaussian process predictions, by using only basis functions surrounding each prediction point to create a local approximation free of boundary effects. We demonstrate the scalability and accuracy of our approach compared to state-of-the-art on data describing natural sound, precipitation levels in the US, and a global bathymetry map.

/* */