CLARION

CLARION
Graeme E. Smith · CC BY-SA 3.0 · source
Name	CLARION
Developer	Ron Sun
First release	1994
Latest release	2010s
Programming language	Common Lisp
Platform	UNIX/Linux, Windows
License	Proprietary/Academic

CLARION

CLARION is a cognitive architecture developed to model human cognition and learning through interacting subsystems. It was proposed to integrate findings from Psychology research such as Dual-process theory, computational models like ACT-R, and neural network approaches exemplified by Backpropagation systems. The project has been used to simulate phenomena studied by researchers affiliated with institutions like University of Toronto, Indiana University, and Rensselaer Polytechnic Institute.

Overview

CLARION posits a dual-layered cognitive system inspired by contrasts between Daniel Kahneman's System 1 and System 2 descriptions, earlier work by Anthony Marcel, and cognitive architectures including SOAR and EPIC. It distinguishes an explicit symbolic subsystem influenced by Noam Chomsky's ideas and an implicit subsymbolic subsystem reminiscent of David Rumelhart's connectionist frameworks. The architecture aims to account for phenomena investigated in experiments such as the Stroop effect, serial recall, and procedural learning paradigms studied by researchers like Endel Tulving and B.F. Skinner.

Architecture and Components

CLARION's structure includes several interacting components analogous to modules in architectures like ACT-R and SOAR: a top-level explicit cognitive module that uses symbolic rules similar to John Anderson's productions, and a bottom-level implicit module that uses distributed representations akin to Geoffrey Hinton's deep networks. Its metacognitive subsystem shares aims with work from Marvin Minsky and Allen Newell on control structures. Other named components parallel systems in LIDA and Global Workspace Theory implementations researched at University of California, San Diego and MIT. The interface between explicit and implicit layers draws on studies by Michael Posner and computational techniques used in Hidden Markov Model applications by L. Rabiner.

Learning Mechanisms

Learning in CLARION combines symbolic rule induction similar to Inductive Logic Programming methods used in Ross Quinlan's algorithms and subsymbolic adaptation via variations of backpropagation as in David Rumelhart and Yann LeCun's work. Reinforcement learning elements reflect approaches pioneered by Richard Sutton and Andrew Barto, and incorporate ideas from Temporal Difference learning studies such as by Gerald Tesauro. Skill acquisition modeled in CLARION parallels research on proceduralization by Herbert Simon and Allen Newell and experimental paradigms explored by Robert Bjork. Meta-learning control mirrors procedures found in Kolodner's Case-based reasoning literature and heuristics from Patrick Winston.

Applications and Implementations

CLARION has been applied to domains ranging from psycholinguistics simulations of sentence comprehension and language acquisition akin to work by Noam Chomsky and Jean Berko Gleason, to modeling moral decision-making experiments related to studies by Joshua Greene and Jonathan Haidt. Implementations have targeted human factors tasks studied at NASA and cognitive modeling related to air traffic control scenarios used by MITRE Corporation researchers. The architecture has been embedded in robotic platforms similar to systems at Carnegie Mellon University and Stanford University to explore embodied cognition themes examined by Rodney Brooks and Hod Lipson. CLARION has also contributed to computational models in social cognition experiments researched at Harvard University and Princeton University.

Performance and Evaluation

Evaluations of CLARION compare its predictive accuracy to architectures like ACT-R, SOAR, and connectionist models evaluated in benchmarks from Cognitive Science Society proceedings and competitions such as those organized by DARPA. Performance metrics often derive from experimental datasets used in studies by Elizabeth Loftus on memory, Daniel Kahneman on decision-making, and Eleanor Rosch on categorization. Comparative analyses consider fit to human behavioral data in tasks including probabilistic reasoning experiments from Amos Tversky and Daniel Kahneman, and skill learning timelines reported by Fitts and Posner.

Criticism and Limitations

Critiques of CLARION focus on issues familiar from debates involving Connectionism versus Symbolic AI proponents such as Jerry Fodor and Zenon Pylyshyn, including questions about scalability raised in work by Judea Pearl and tractability concerns discussed by Leslie Valiant. Empirical challenges mirror criticisms leveled at other architectures in publications by Gary Marcus regarding systematicity and compositionality, and methodological debates found in Replication crisis literature highlighted by Open Science Collaboration. Practical limitations include integration difficulties noted in interdisciplinary reviews from Cognitive Science Society and engineering constraints discussed in reports by Defense Advanced Research Projects Agency.

Category:Cognitive architectures