Interactive demo · Probabilistic Machine Learning: Advanced Topics · Chapter 4

Gaussian Bayes nets and Gaussian MRFs

A linear-Gaussian DAG, xi = ∑j wij xj + si zi (4.12), is a Gaussian with covariance Σ = US2UT, U = (I − W)−1 (4.16), and precision Λ = (I − W)TS−2(I − W). The precision is sparse: its zero pattern is the moralised graph, and Λij = 0 exactly when xi and xj are conditionally independent given all the others. The covariance is usually dense. Change the weights and the noise level, click a cell to compare the marginal and the conditional dependence of a pair.

Graph

Edge weight scale (all weights times a)

Conditional std s (all nodes)

Moral edges in the graph

Try these

Nodes are numbered in a topological order and all means are zero (they do not change Σ or Λ). An entry is called zero when its absolute value is below 10−9. For the AR(1) chain the first node starts in the stationary distribution, s1 = s/√(1 − a²w²), as in (4.106).

In the lecture notes