Scalable backpropagation for Gaussian Processes using celerite
read the original abstract
This research note presents a derivation and implementation of efficient and scalable gradient computations using the celerite algorithm for Gaussian Process (GP) modeling. The algorithms are derived in a "reverse accumulation" or "backpropagation" framework and they can be easily integrated into existing automatic differentiation frameworks to provide a scalable method for evaluating the gradients of the GP likelihood with respect to all input parameters. The algorithm derived in this note uses less memory and is more efficient than versions using automatic differentiation and the computational cost scales linearly with the number of data points.
This paper has not been read by Pith yet.
Forward citations
Cited by 3 Pith papers
-
A Model Selection Criterion for Multidimensional Gaussian Processes: Application to Radial Velocities
Introduces MGIC_rv, an information criterion that combines conditional RV likelihood with an effective parameter count for selecting multi-GP models focused on radial velocities.
-
The X-ray emission of the long-period transient and accreting cataclysmic variable ASKAP J174508.9-505149
X-ray timing and spectral analysis of ASKAP J174508.9-505149 detects matching periodicity and features consistent with an accreting magnetic CV.
-
The mass of TOI-1883 b: A low density super-Neptune in the ridge regime transiting an early-M dwarf
Mass of 13.7 Earth masses and density 0.4 g cm^{-3} measured for TOI-1883 b, a super-Neptune in the ridge regime around an early-M dwarf, with implications for disk migration and photoevaporation.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.