Interactive demo · Bayesian and Probabilistic Machine Learning · Bridge to deep generative models

Sequential Bayesian linear regression

Each point updates a Gaussian belief about the intercept w0 and slope w1 of a line. Left: the posterior density over (w0, w1), with the true values marked by a cross. Right: ten lines drawn from the current posterior, the data, and the true line. Changing the prior precision or the assumed noise precision replays the same data under the new assumptions.

Prior precision α 2.0

2.0

Assumed noise precision β 25

25

The data are generated with true noise sd 0.2 (precision 25) in every case, so a small β expects noisier data than there is.

Stream data

Scoreboard

Try these

This is Bayes' rule one point at a time: the posterior after n points is the prior times the likelihood of all n points, so the order of the points does not matter (Bishop eqs. 3.49 to 3.51, and the batch form 3.53 to 3.54).

In the lecture notes