Paul Rosenbloom: “Towards Functionally Elegant, Grand Unified Architectures”

March 14, 2012 | Amelia Island, Florida

Speaker: Paul Rosenbloom
Host: 21st Annual Conference on Behavior Representation in Modeling Simulation (BRIMS 2012)

When developing cognitive architectures, the ultimate goal is typically a unified theory of intelligent behavior, with the working focus then being on integrating across the capabilities implicated within central cognition, and the result being a unified architecture for cognition. What can be called a grand unified architecture sets the bar higher, striving to also include the key non-cognitive aspects of intelligent behavior, such as perception, motor control, personality, motivation and affect. Such architectures can further be considered functionally elegant if they provide the requisite breadth of functionality in a simple and theoretically elegant manner, yielding a form of cognitive Newton’s laws that provides broad coverage from interactions among a small set of general principles/mechanisms. The pursuit of functionally elegant, grand unified architectures provides a challenging research path, yet one that points the way towards rapid progress beyond today’s state of the art, even within the more traditional cognitive focus; and which should also support both deep science and useful systems. I am currently approaching this goal by rethinking architectures from the ground up, leveraging the interactions between a pair of very general mechanisms – graphical models (factor graphs, in particular) and piecewise continuous multivariate functions – to yield a parameterized space of state-of-the-art capabilities over the processing of symbols, probabilities and signals. The availability of this broad parameterized space promises to accelerate the evolution of cognitive architectures by facilitating the exploration of a wider range of the requisite capabilities and their variations; and without the need to explicitly implement a whole new module for each. Work to date – much of which will be summarized here – demonstrates that within the resulting space can be found: standard flavors of long-term memory, such as a procedural rule-based memory and declarative semantic and episodic memories, plus other variations and blends; forms of knowledge-based, decision-theoretic and social problem solving; perception and mental imagery; and key bits of language processing. Much more is still required on many of these topics, and additional capabilities must also be added, but the already proven applicability of graphical models to many of these problems shines a bright light on the path towards their rapid incorporation into such a grand synthesis. It also may help understand other topics – such as personality, motivation and affect – that have not previously been investigated via these kinds of techniques. For some capabilities – such as learning – more principles/mechanisms will likely be required, but functional elegance still looks to be within reach, with the inclusion of only a small number of additional general principles/mechanisms.