Exploiting Graphical Models for Enhancing Functionality and Uniformity in Cognitive Architectures

Paul Rosenbloom
University of Southern California (USC)

A cognitive architecture, whether intended as a model of human cognition or as the basis for artificially intelligent (AI) systems, provides a hypothesis about the fixed mechanisms underlying intelligence and how they combine to yield appropriate environmental behavior in the presence of knowledge. The ideal architecture combines a high level of functionality (the range of intelligent behaviors that are realizable) and uniformity (producing such behaviors from interactions among a small number of general mechanisms) to yield both practical utility and theoretical elegance. However, existing architectures fall short of this ideal along both of these dimensions, and further often find these dimensions at variance with each other. In this talk I will discuss an approach to building more functional hybrid (discrete and continuous) mixed (Boolean and Bayesian) cognitive architectures based on the uniformity of signal, probability and symbol processing enabled by graphical models. Progress to date will be presented on implementing a hybrid mixed memory architecture – combining a rule-based procedural memory, semantic and episodic declarative memories, and a constraint memory – uniformly via factor graphs and the summary product algorithm.

Presentation (PowerPoint File)

Back to Machine Reasoning Workshops I & II: Mission-Focused Representation & Understanding of Complex Real-World Data