GPflow

Why we included this project

Gaussian processes return a prediction and an honest measure of how much to trust it, and GPflow is a Python library for building those models with TensorFlow handling the heavy computation underneath. That uncertainty is why it keeps showing up in Bayesian optimisation and experiment design: the posterior tells you where to sample next instead of just pushing out another guess. Kernels and likelihoods compose cleanly, so you can assemble a custom model without hand-writing inference, and variational methods keep things tractable as datasets grow. Researchers use it to prototype new GP formulations and still train on moderately large data, and the documentation plus an active Slack community make it a reasonable starting point whether you are new to Gaussian processes or porting an existing implementation.