Perceptual Inference Network (PIN): Formulation
Boyer - 21
• Bayesian networks are directed acyclic graphs with nodes repre
senting propositions (or random variables) and arcs signifying di
rect dependencies as quantified by conditional probabilities. The
conditional dependencies are made explicit by the use of a graph
ical structure.
• Our Problem: We have m features, / 1 , • • •, / m , at n locations
/ 1 , • • •, / n . We need a means of efficient updating of probabilities
of these features. Let us form our binary random variables as lj 3 .
• Observations:
1. Dependent features tend to be close together spatially.
2. Similar feature types at different location will have similar
neighborhood structure.
• Solution: Group the random variables , • • •, l[ 3 to form a com
posite random variable representing the events that feature fj
occurs at lj through /*.