This dev-meetup is split in several smaller notebooks to keep each one of the notebooks focused.
- From Linear Regression to Gaussian Processes explains the weight-space view of Gaussian Processes. It also explains what a kernel or covariance function is and how you arrive at this concept relatively naturally once you introduce projections of inputs into feature space.
- Sampling from Gaussian Process by Hand shows how to sample functions from a gaussian process in a similar way like you would sample numbers from a normal probability distribution.
- Gaussian Process Parameter Effects shows what effects the modification of the hyper parameters of the convariance funcion (a.k.a. kernel) has on the recovered regression line.
- CO2 Mauna Loa Gaussian Process Regression shows how to perform the decomposition of additive effects to explain parts of the overall model.
- The Birthdates demo using Gaussian Processes from the matlap GPstuff package is even more impressive. It is also the picture on the front cover of Bayesian Data Analysis, Third Edition book.
- In the blog post Laurie Davies: time series decomposition of birthday data links to more recent analysis are given.
- David Kristjanson Duvenaud goes even one step further in his phd-thesis Automatic Model Construction with Gaussian Processes. He explains how to express structure with kernels, how to automate the model building process and even how to convert such a model automatically to English language as a kind of report. This is part of the Automatic Statistician project by among others, Zoubin Ghahramani, Uber's Chief Scientist.