Skip to content

Accompanying tutorials for my talk "Kriging in the age of machine learning", presented at the UCL-MOAP workshop on "Bayesian Machine Learning for Weather and Climate".

Notifications You must be signed in to change notification settings

UCL-SML/moap-bml-workshop

 
 

Repository files navigation

Gaussian Process Tutorials for UCL-MOAP Workshop

This repo contains the tutorials that I made in companion to my presentation at the UCL-MOAP workshop on "Bayesian Machine Learning for Weather and Climate". In particular, you can find implementations of all the experiments shown in the presentation and further details on various techniques that I didn't have time to present at the workshop.

The tutorial is divided into 5 parts, which cover the following topics.

  1. Basics of Gaussian process regression including how to set up a model, how to condition on data and how to optimise for the hyperparameters using the log marginal likelihood.
  2. Sparse GPs, in particular Titsisas' inducing point method. Includes information on the mathematical derivation, comparisons with a vanilla Gaussian process and training.
  3. Gaussian process inference for timeseries data using Kalman filtering or smoothing techniques.
  4. Global temperature interpolation using Matérn GPs defined on the sphere.
  5. Global wind speed interpolation using projected kernels.

gp_conditioning_vid

Dependencies

To run the first three notebooks, please install the packages in the requirements.txt file by running

pip install -r requirements.txt

The codes in the fourth and fifth tutorials use our riemannianvectorgp package, which can be found here. This is not on pip so you have to clone this into your repository and install the requirements therein. To plot the results, I used the cartopy package, which is easiest to install using conda by running

conda install -c conda-forge cartopy

Data

All the datasets used in the tutorials are saved in the data directory.

About

Accompanying tutorials for my talk "Kriging in the age of machine learning", presented at the UCL-MOAP workshop on "Bayesian Machine Learning for Weather and Climate".

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 99.9%
  • Python 0.1%