Interactive demo · Foundations of Deep Learning · Chapter 3

The multivariate Gaussian: conditioning, marginalising and Bayes' theorem

A two-dimensional Gaussian p(xa, xb) is drawn as its 1 and 2 standard-deviation ellipses, with the eigenvectors of Σ scaled by λi1/2. Fix xb and the slice through the joint density, renormalised, is the conditional p(xa | xb) (red); integrating xb out gives the marginal p(xa) (blue). The second mode reads the same picture as a linear-Gaussian model (prior, observation, posterior); the third draws data and fits it by maximum likelihood.

Mode

Covariance

Try these

All numbers are exact formulas except in the maximum likelihood mode, where data are drawn with a seeded generator (the same numbers as the Python code).

In the lecture notes