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

Inference and d-separation in a Bayes net

Click a state of any node (or use the buttons) to observe it; every other node shows its exact posterior marginal, computed by summing the joint distribution (4.2). Observing a child of a collider makes its parents dependent: the student's grade of C is explained either by a hard class or by low intelligence, and a good SAT score then removes the second explanation (explaining away, Figure 4.8). The tester below asks whether two nodes are d-separated by the observed set, and checks the answer against the exact conditional mutual information. The last network is the Gaussian collider of Figure 4.6, where selecting on the child z correlates the independent parents.

Network

Evidence (observe a state)

Are these d-separated?

and
given the set of observed nodes (the values observed do not matter).

Try these

Posterior marginals are exact (enumeration of the joint). The student CPTs are those of Figure 4.2; the sprinkler CPT of W is the one printed in the supplement and the other tables are the standard ones that reproduce its quoted probabilities. The Gaussian panel uses 6000 seeded samples; the correlation quoted is the exact one, from the truncated normal.

In the lecture notes