Sum-product networks (SPNs) are a new kind of deep architecture that support exact, tractable inference over a large class of problems for which traditional graphical models cannot. The Sigma cognitive architecture is based on graphical models, posing a challenge for it to handle problems within this class, such as parsing with probabilistic grammars, a potentially important aspect of language processing. This work proves that an early unidirectional extension to Sigma’s graphical architecture, originally added in service of rule-like behavior but later also shown to support neural networks, can be leveraged to yield exact, tractable computations across this class of problems, and further demonstrates this tractability experimentally for probabilistic parsing. It thus shows that Sigma is able to specify any valid SPN and, despite its grounding in graphical models, retain the desirable inference properties of SPNs when solving them.
