Sigma (Σ) graphical Cognitive Architecture is developed by a research group under Paul S. Rosenbloom - Project Leader in the Institute for Creative Technologies (ICT), Playa Vista, CA, USA at the University of Southern California (USC) and a Professor in the Computer Science Department, Viterbi School of Engineering (USC), Los Angeles, CA, USA. Sponsors: the U.S. Air Force Office of Scientific Research and the U.S. Army.
The Sigma (Σ) architecture represents an attempt to rethink cognitive architectures from the ground up via graphical models. It has been under development in some form since 2008, under the general rubric of a graphical cognitive architecture, but due to ambivalence as to whether what was being built was a specific architecture or an environment in which to explore a range of architectures, it was not until the summer of 2012 – by when it was overwhelmingly obvious that at least one specific architecture was being built – that it received a name. Sigma is intended to be a proper noun rather than an acronym, although it was inspired by “Sigma is an integrated (or intelligent) graphical model architecture”.
Developing Sigma (Σ), an attempt to build a functionally elegant, grand unified cognitive architecture/system – based on graphical models and piecewise continuous functions – in support of virtual humans and intelligent agents/robots.
Work to date has focused on memory and learning; decision making and problem solving; reflection and Theory of Mind; perception, localization and mental imagery; language and speech. More on this effort can be found here.
The Sigma (Σ) architecture is built to be hybrid from the ground up, in service of satisfying two general desiderata: grand unification and functional elegance.
A traditional unified cognitive architecture attempts to bring together in an integrated manner the range of cognitive capabilities required for human(-level) intelligent behavior in the world. A grand unified architecture goes beyond this, in analogy to a grand unified theory in physics, to attempt to include the crucial pieces missing from a purely cognitive theory, such as perception, motor control, and emotion.
Functional elegance implies a combination of the broad range of capabilities required in a (grand) unified architecture with simplicity and theoretical elegance.
In Sigma, the aim is something like a set of cognitive Newton’s laws that yield the required diversity of behavior from interactions among a small set of very general primitives. Within AGI, AIXI  can be seen as an attempt at an extreme form of functional elegance. The approach in Sigma is less ambitious, but still strongly in this direction.
Driven by these desiderata, work to date on Sigma has been deliberately broad – including forms of memory and learning, problem solving and decision making, perception and localization, and natural language – with the intent of determining whether a small set of general mechanisms can in fact be sufficient in combination.
Thus, for mental imagery the natural question to ask became whether Sigma could provide a sufficient hybrid capacity without either distinct symbolic versus imagistic modules or distinct representations, memories and processes, as has been necessary in other architectural approaches.
Mental imagery in Sigma is grounded in:
(1) the architecture’s generalized language of conditionals, which compiles down to factor graphs for processing via the summary product algorithm;
(2) an inherently continuous piecewise linear representation for the functions and messages in (1); and
(3) affine transformations – a generalization of the offsets introduced earlier – and piecewise linear filters.
By demonstrating mental imagery via interactions among more primitive mechanisms, this work contributes to the breadth of functionality unified within Sigma, while doing so in a simple and elegant manner.
The key to functional elegance here has been to begin with a small set of very general mechanisms that are leveraged in combination when possible, and which are (minimally) augmented when necessary. This combination also supports grand unification, intertwining continuous perception-related information with general symbol processing.
Rosenbloom, P.S.: Deconstructing reinforcement learning in Sigma. In: Fifth Conference on Artificial General Intelligence (2012)
Rosenbloom, P.S.: Graphical models for integrated intelligent robot architectures. In: AAAI Spring Symposium on Designing Intelligent Robots (2012)
Rosenbloom, P.S.: From memory to problem solving: Mechanism reuse in a graphical cognitive architecture. In: Fourth Conference on Artificial General Intelligence (2011)
Chen, J., Demski, A., Han, T., Morency, L-P., Pynadath, P., Rafidi, N., Rosenbloom, P.S.: Fusing symbolic and decision-theoretic problem solving + perception in a graphical cognitive architecture. In: Second International Conference on Biologically Inspired Cognitive Architectures (2011)
Rosenbloom, P.S.: Mental imagery in a graphical cognitive architecture. In: Second International Conference on Biologically Inspired Cognitive Architectures (2011)
Rosenbloom, P.S.: Bridging dichotomies in cognitive architectures for virtual humans. In: AAAI Fall Symposium on Advances in Cognitive Systems (2011)
Rosenbloom, P.S.: Implementing first-order variables in a graphical cognitive architecture. In: First International Conference on Biologically Inspired Cognitive Architectures (2010)
Rosenbloom, P.S.: Combining Procedural and Declarative Knowledge in a Graphical Architecture. In: 10th International Conference on Cognitive Modeling (2010)