Baseline¶

Algorithms:

• Inverse Weighted Distances

• Radial Basis Function

• Window Function

• Universal Kriging

Gaussian Processes¶

• Scratch: Gaussian Processes + Jax + Cola

• Library: GPJax + Cola

• Custom Library: TinyGP + Cola + Custom Solver

Scaling I - Hardware¶

• Scratch: KeOPs + Gaussian Processes

• Library: KeOPs + GPyTorch

• Library: KeOps + Nystrom

Scaling II - Algorithm¶

Here, we try to take advantage of algorithm speed-ups

Model Approximations¶

• RBFSampler/Nystrom + Linear SGD

• KeOps GPU + Gaussian Kernel

Sparse Approximations¶

These methods are some speed-ups that can be achieved from using SOTA Gaussian processes.

• Sparse Gaussian Process (SGP)

• Stochastic Variational Gaussian Processes (SVGP)

• KISS-GP

• Deep Kernel Learning

Basis Functions¶

• Deep Kernel Learning

• RFF Gaussian Processes

• Spherical Gaussian Process

Dynamical¶

• Linear Kalman Filter

• Non-Linear Kalman Filter

• Markovian Gaussian Processes

Engineering Tricks¶

Feature Engineering¶

Scaling - MinMaxScaler, StandardScaler

Temporal Features - Coordinates, Cyclic, Splines, Fourier Features, Sinusoidal Features

Spatial Features - Coordinate Transforms (Cartesian, Spherical, Cylindrical), Cycle, Splines, Fourier Features, Spherical Harmonics

Patches vs Radius¶