Speaker: Guillaume Rongier

Date: Thursday 13th of July 2023, 1:15pm.

Abstract:

Recent developments in deep learning have turned the spotlight on the entire field of machine learning. This has led to more and more studies comparing predictions from geostatistical and machine learning methods, sparking debates about whether machine learning will take over geostatistics in geological applications. This implicitly implies an opposition between those two fields. In this context, a statement made by Williams in 1998 contains an intriguing perspective: "In the Bayesian approach to neural networks, a prior on the weights of a network induces a prior over functions. An alternative method of putting a prior over functions is to use a Gaussian process (GP) prior over functions. This idea has been used for a long time in the spatial statistics community under the name of 'kriging' although it seems to have been largely ignored as a general purpose regression method." If the method known as kriging in geostatistics is the same as the method known as Gaussian processes in machine learning, how is it that geostatistics and machine learning are so often introduced as contending fields? Can we find similar relationships with more recent geostatistical approaches such as multiple-point simulation? What does this mean for the future development of both fields? This talk will deconstruct the idea that geostatistics and machine learning are two completely separate fields: their methods share a lot more similarities than usually acknowledged, and geostatistics can be seen as a sub-field of machine learning specialized on predictions with spatial data. Since 1998, Gaussian processes have become an integral part of machine learning, and both fields stand to gain from further exploring their similarities.