Equivariant Studying of Stochastic Fields: Gaussian Processes and Steerable Conditional Neural Processes
Authors: Peter Holderrieth, Michael Hutchinson, Yee Whye Teh
Summary: Motivated by objects akin to electrical fields or fluid streams, we research the issue of studying stochastic fields, i.e. stochastic processes whose samples are fields like these occurring in physics and engineering. Contemplating basic transformations akin to rotations and reflections, we present that spatial invariance of stochastic fields requires an inference mannequin to be equivariant. Leveraging latest advances from the equivariance literature, we research equivariance in two courses of fashions. Firstly, we totally characterise equivariant Gaussian processes. Secondly, we introduce Steerable Conditional Neural Processes (SteerCNPs), a brand new, totally equivariant member of the Neural Course of household. In experiments with Gaussian course of vector fields, photographs, and real-world climate information, we observe that SteerCNPs considerably enhance the efficiency of earlier fashions and equivariance results in enhancements in switch studying duties
