Interactive demo · Foundations of Deep Learning · Chapter 11

Explaining away and d-separation, with exact probabilities

A directed graph says how a joint distribution factorises, p(x) = ∏k p(xk | pak), and therefore which conditional independences hold. Click a node (or use the buttons) to observe it as 0 or 1. Every probability on the page is computed exactly by summing the joint over the unobserved nodes. Watch a head-to-head node: its parents are independent until it, or one of its descendants, is observed; then evidence for one cause explains away the other.

Network

Observe (? = not observed)

Is A ⊥ B | observed nodes?

AB

Conditional probability tables

Try these

In the lecture notes