Lookup Table

Table with 2^k rows can specify an arbitrary CPT, where k is number of parents of a node.

However, this is too complicated

Deterministic Node

“AND” gate will check whether all of them are 1,
“OR” gate will check whether all of them are 0.

However, this is too simple

Noisy-OR

Use k numbers to parametrize all 2^k entries.

More parents are equal to 1, the more probable that Y=1.

However, this only takes into account excitation!

Sigmoid CPT

Use k real numbers to parametrize 2^k elements in CPT

Also known as logistic regression in statistics,
activation function in neural networks.

Different from Noisy-OR in that this can mix both inhibition and excitation!