Comparative Table of Cognitive Architectures

Started on October 27, 2009 | last update: June 18, 2012 | Curator: Alexei Samsonovich

Features vs. Architectures 4D/RCSACT-RARTBECCAbiSoarCERA-CRANIUMChrestClarionCogPrimeCoJACKDiscipleEpicFORRGLAIRGMU-BICAHTMLeabraLIDANARSNextingPogamutPolyschemeRecommendation ArchitectureREMSoarYmirother architectures to consider including:
Contributors to this table James AlbusAndrea Stocco and Christian LebiereStephen GrossbergBrandon Rohrer, updated 8/3/2011B. Chandrasekaran and U. KurupRaul ArrabalesFernand Gobet and Peter LaneRon SunBen GoertzelFrank Ritter & Rick EvertszGeorge TecuciShane Mueller, Andrea StoccoSusan L. EspteinStuart C. ShapiroAlexei SamsonovichJeff HawkinsDavid C. NoelleStan FranklinPei WangAkshay Vashist, Shoshana LoebCyril BromNick CassimatisL. Andrew CowardAshok K. Goel, J. William Murdock, and Spencer RugaberAndrea Stocco / John LairdKristinn R. Th_risson
Iconic link to diagramsIcarus SAL Tosca
Knowledge and experiences are represented usingimages, maps, objects, events, state, attributes, relationships, situations, episodes, framesChunks and productions,Visual 3D boundary and surface representations; auditory streams; spatial, object, and verbal working memories; list chunks; drive representations for reinforcement learning; orienting system; expectation filter; spectral timing networksPerceptual experiences are represented features and combinations of features. Knowledge about the world is represented as transitions between experiences.The representational framework of Soar plus diagrams ? the diagrammatic part can also be combined with any symbolic general architecture, such as ACT-RSingle percepts, complex percepts, and mission percepts. Single and complex behaviors.chunks and productionschunks, rules, and NNsCogPrime is a multi-representational system. The core representation consists of a hypergraphs with uncertain logical relationships and associative relations operating together. Procedures are stored as functional programs; episodes are stored in part as ?movies? in a simulation engine; and there are other specialized methods too.Beliefs-Desires-Intentions (BDI) architecture with events, plans/intentions (procedural), beliefsets (declarative) and activation levelsDisciple shell includes general modules for user-agent interaction, ontology representation, problem solving, learning, and tutoring, as well as domain-independent knowledge (e.g., knowledge for evidence-based reasoning). The problem solving engine of a Disciple cognitive assistant (see the top part of Figure 6) employs a general divide-and-conquer strategy, where complex problems are successively reduced to simpler problems, and the solutions of the simpler problems are successively combined into the solutions of the corresponding complex problems. To exhibit this type of behavior, the knowledge base of the agent contains a hierarchy of ontologies, as well as problem reduction rules and solution synthesis rules which are expressed with the concepts from the ontologies.Production Rules, Working memory entries Descriptives are shared knowledge resources computed on demand and refreshed only when necessary Advisors are domain-dependent decision rationales for actions. Measurements are synopses of problem solving experiences. SNePS, simultaneously a logic-based, assertional frame-based, and propositional graph based representationschemas, mental statesSparse distributed representations. Memory and representations are distributed across a hierarchy of nodes. Within each node representations are large sparse binary vectors.patterns of neural firing rates and patterns of synaptic strengths. Sensory events drive patterns of neural activation, and such activation-based representations may drive further processing and the production of actions. Knowledge that is retained for long periods is encoded in patterns of synaptic connections, with synaptic strengths determining the activation patterns that arise when knowledge or previous experiences are to be employed.Perceptual knowledge - nodes and links in a Slipnet-like net with sensory data of various types attached to nodes; Episodic knowledge - boolean vectors (Sparce Distributed Memory; Procedural knowledge - schemes a la Schema Mechanismbeliefs, tasks, and conceptsFacts, Rules, frames (learned/ declared), symbolic representation of raw sensory inputs, expectation generation and matchingrules, operatorsRelational constraints, constraint graphs, first-order literals, taxonomies, weight matrixes.A large set of heuristically defined similarity circumstances, each of which is a group of information conditions that are similar and have tended to occur at the same time in past experience. One similarity circumstance does not correlate unambiguously with any one cognitive category, but each similarity circumstance is associated with a range of recommendation weights in favour of different behaviours (such as identifying categories in current experience). The predominant weight across all currently detected similarity circumstances is the accepted behaviourTasks, methods, assertions, tracesProcedural knowledge: Rules. Semantic knowledge: relational graph structure, Episodic memory: episodes of relational graph structuresDistributed modules with traditional Frames4CAPS AIS Apex Atlantis CogNet Copycat DUAL Emotion Machine
Main componentsBehavior Generation, World Modeling, Value Judgment, Sensory Processing, Knowledge Database. These are organized in a hierarchical real-time control system (RCS) architectureChunk-based declarative memory; buffers; procedural knowledge encoded as productionsMany model brain regions, notably laminar cortical and thalamic circuitsBECCA is a solution to the general reinforcement learning problem. It consists of two parts, an unsupervised feature creator and a model-based incremental learner. Both are incremental and on-line, designed for a physically embodied agent operating in an unstructured environment.DRS, for diagram representation; perceptual and action routines to get information from and create/modify diagrams.Physical layer global workspace, Mission-specific layer global workspace, core layer contextualization, attention, sensory prediction, status assessment, goal management.Attention, sensory memory, short-term memory, long-term memoryprocedural knowledge and declarative knowledge (each in both implicit and explicit forms --- rules and NNs)The primary knowledge store is the AtomSpace, a neural-symbolic ?weighted labeled hypergraph? with multiple cognitive processes acting on it (in a manner carefully designed to manifest cross-process ?cognitive synergy?), and other specialized knowledge stores indexed by it. The cognitive processes are numerous but include: an uncertain inference engine (PLN, Probabilistic Logic Networks), a probabilistic evolutionary program learning engine (MOSES, developed initially by Moshe Looks), an attention allocation algorithm (ECAN, Economic Attention Networks, which is somewhat neural net like), concept formation and blending heuristics, etc. Work is under way to incorporate a variant of Itamar Arel?s DeSTIN system as a perception and action layer. Motivation and emotion are handled via a variant of Joscha Bach?s MicroPsi framework called CogPsi.Beliefsets for long term memory, plans, intentions, events, goals Cognitive processor (including production rule interpreter and working memory), long term memory, production memory, detailed perceptual-motor interfaces (auditory processor, visual processor, ocular motor processor, vocal motor processor, manual motor processor, tactile processor)Advisors are organized into three tiers. Tier-1 Advisors are fast and correct, recommend individual actions, and are consulted in a pre-specified order. Tier-2 Advisors trigger in the presence of a recognized situation, recommend (possible partially ordered) sets of actions, and are consulted in a pre-specified order. Tier-3 Advisors are heuristics, recommend individual actions, and are consulted together. Tier-3 Advisors' opinions express preference strengths that are combines with weights during voting to select an action. A FORR-based system learns those weights from traces of its problem-solving behavior.1) Knowledge Layer containing Semantic Memory, Episodic Memory, Quantified & conditional beliefs, Plans for non-primitive acts, Plans to achieve goals, Beliefs about preconditions & effects of acts, Policies (Conditions for performing acts), Self-knowledge, Meta-knowledge; 2) Perceptuo-Motor Layer containing implementations of primitive actions, perceptual structures that ground KL symbols, deictic and modality registers; 3) Sensori-Actuator Layer containing sensor and effector controllers.memory systems: working, semantic, episodic, procedural, iconic (I/O); plus: cognitive map, reward system, the engineHTM is a biologically constrained model of neocortex and thalamus. HTM models cortex related to sensory perception, learning to infer and predict from high dimensional sensory data. The model starts with a hierarchy of memory nodes. Each node learns to pool spatial pattern using temporal contiguity (uising variable order sequences if appropriate) as the teacher. HTMs are inherently modality independent. Biologically the model maps to cortical regions, layers of cells, columns of cell across the layers, inhibitory cells, and non-linear dendrite properties. All representations are large sparse distributions of cell activities.At the level of gross functional anatomy, most Leabra models employ a tripartite view of brain organization. The brain is coarsely divided into prefrontal cortex, the hippocampus and associated medial-temporal areas, and the rest of cortex -- "posterior" areas. Prefrontal cortex provides mechanisms for the flexible retention and manipulation of activation-based representations, playing an important role in working memory and cognitive control. The hippocampus supports the rapid weight-based learning of sparse conjunctive representations, providing central mechanisms for episodic memory. The posterior cortex mostly utilizes slow statistical learning to shape more automatized cognitive processes, including sensory-motor coordination, semantic memory, and the bulk of language processing. At a finer level of detail, other "components" regularly appear in Leabra-based models. Activation-based processing depends on attractor dynamics utilizing bidirectional excitation between brain regions. Fast pooled lateral inhibition plays a critical role in shaping neural representations. Learning arises from an associational "Hebbian" component, a biologically plausible error-driven learning component, and a reinforcement learning mechanism dependent on the brain's dopamine system.Cognitive cycle (action-perception cycle) acting as a cognitive atom. Higher level processes implemented as behavior streams. Cognitive cycle includes sensory memory, perceptual associative memory, workspace, transient episodic memory, declarative memory, global workspace, procedural memory, action selection, sensory motor memoryan inference engine and an integrated memory and control mechanismLearning, Reasoning, Imagining, Attention Focus, Time awareness, Expectation generation and matching procedural knowledge encoded as reactive rules; episodic and spatial memory encoded within a set of graph-based structuresModules using data structres and algorithms specialized for specific concepts. A focus of attention for exchaning information among these modules. A focus manager for guiding the flow of attention and thus inference. Specialized modules for representing and making inferences about specific concepts.Condition definition and detection (cortex); Selection of similarity circumstances to be changed in each experience (hippocampus); Selection of sensory and other information to be used for current similarity circumstance detection (thalamus); Assignment and comparison of recommendation weights to determine current behaviour (basal ganglia); Reward management to change recommendation weights (nucleus accumbens etc.); Management of relative priority of different types of behaviour (amygdala and hypothalamus); Recording and implementation of frequently required behaviour sequences (cerebellum)Task-Method-Knowledge (TMK) models provide functional models of what agents know and how they operate. They describe components of a reasoning process in terms of intended effects, incidental effects, and decomposition into lower-level components. Tasks include the requirements and intended effects of some computation. Methods implement a task and include a state-transition machine in which transitions are accomplished by subtasks.Working memory encoded as a graph structure; Knowledge represented as rules organized as operators; semantic memory; episodic memory; menal imagery; reinforcement learningDistributed heterogeneous module-based network, interacting via blackboards, realtime performanceERE Gat Guardian H-Cogaff Homer Imprint MAX Omar PRODIGY PRS Psi-Theory R-CAST RALPH-MEA Society of Mind Subsumption architecture Teton Theo
Primary original reference(s)Anderson, J.R., & Lebiere, C. (1998). The Atomic Components of Thought. Mahwah: Lawrence Erlbaum Associates.Grossberg, S. (1987). Competitive learning: From interactive activation to adaptive resonance. Cognitive Science 11: 23-63.B Rohrer, "An implemented architecture for feature creation and general reinforcement learning," Workshop on Self-Programming in AGI Systems, Fourth International Conference on Artificial General Intelligence, Mountain View, CA, Aug 3-6, 2011. (1MB pdf) http://www.sandia.gov/rohrer/doc/Rohrer11ImplementedArchitectureFeature.pdfUnmesh Kurup & B. Chandrasekaran, "Modeling Memories of Large-scale Space Using a Bimodal Cognitive Architecture." Proceedings of the International Conference on Cognitive Modeling July 27-29, 2007, Ann Arbor, MI. (CD-ROM, 6 pages.)Arrabales, R. Ledezma, A. and Sanchis, A. "A Cognitive Approach to Multimodal Attention". Journal of Physical Agents. Volume 3. Issue 1. Pages 53-64. January 2009. Feigenbaum, E. A., & Simon, H. A. (1962). A theory of the serial position effect. British Journal of Psychology, 53, 307-320. Gobet, F. & Lane, P. (2005). The CHREST architecture of cognition: Listening to empirical data. In D. Davis (Ed.), Visions of mind: Architectures for cognition and affect. (pp. 204-224). Hershey, PA: Information Science Publishing.Sun, R. (2004). The CLARION cognitive architecture: Extending cognitive modeling to social simulation. In: Ron Sun (Ed.), Cognition and Multi-Agent Interaction. Cambridge University Press: New York.The most representative early paper on Disciple is [74]. Other key references include [75-78]. Most recent representative publications include [79,80]. [74] Tecuci G., Disciple: A Theory, Methodology and System for Learning Expert Knowledge, Th_se de Docteur en Science, University of Paris-South, 1988. [75] Tecuci G., Building Intelligent Agents: An Apprenticeship Multistrategy Learning Theory, Methodology, Tool and Case Studies, San Diego: Academic Press, 1998. [76] Tecuci G., Boicu M., Bowman M., Marcu D., with a commentary by Burke M., An Innovative Application from the DARPA Knowledge Bases Programs: Rapid Development of a Course of Action Critiquer, AI Magazine, 22, 2, pp. 43-61, 2001. [77] Tecuci G., Boicu M., Boicu C., Marcu D., Stanescu B., Barbulescu M., The Disciple-RKF Learning and Reasoning Agent, Computational Intelligence, 21, 4, pp. 462-479, 2005. [78] Tecuci G., Boicu M., and Comello J., Agent-Assisted Center of Gravity Analysis, CD with Disciple-COG and Lecture Notes used in courses at the US Army War College and Air War College, GMU Press,, 2008. [79] Tecuci G., Boicu M., Marcu D., Boicu C., Barbulescu M., Disciple-LTA: Learning, Tutoring and Analytic Assistance, Journal of Intelligence Community Research and Development, 2008. [80] Tecuci G., Schum D.A., Boicu M., Marcu D., Hamilton B., Intelligence Analysis as Agent-Assisted Discovery of Evidence, Hypotheses and Arguments. In: Phillips-Wren, G., Jain, L.C., Nakamatsu, K., Howlett, R.J. (eds.) Advances in Intelligent Decision Technologies, SIST 4, pp. 1-10. Springer-Verlag, Berlin Heidelberg, 2010. Kieras, D. & Meyer, D.E. (1997). An overview of the EPIC architecture for cognition and performance with application to human-computer interaction. Human-Computer Interaction., 12, 391-438. ; Meyer, D.E., & Kieras, D.E. (1997). A computational theory of executive cognitive processes and multiple task performance: Part I. Basic mechanisms. Psychological Review 63: 81-97.Susan L. Epstein. 1994. For the Right Reasons: The FORR Architecture for Learning in a Skill Domain. Cognitive Science, 18(3): 479-511.Samsonovich, A. V. and De Jong, K. A. (2005). Designing a self-aware neuromorphic hybrid. In K.R. Thorisson, H. Vilhjalmsson, and S. Marsela (Eds.). AAAI-05 Workshop on Modular Construction of Human-Like Intelligence: AAAI Technical Report, volume WS-05-0 George D, Hawkins J (2009) Towards a Mathematical Theory of Cortical Micro-circuits. PLoS Comput Biol 5(10). "On Intelligence" Jeff Hawkins with Sandra Blakeslee, 2005O?Reilly, R.C. (1996), ?Biologically Plausible Error-Driven Learning Using Local Activation Differences: The Generalized Recirculation Algorithm?, Neural Computation, 8, 895?938.Franklin, S. and F. G. Patterson, Jr. (2006). The Lida Architecture: Adding New Modes of Learning to an Intelligent, Autonomous, Software Agent. Integrated Design and Process Technology, IDPT-2006, San Diego, CA, Society for Design and Process Science. Akshay Vashist, Shoshana Loeb (2010), Attention Focusing Model for Nexting Based on Learning and Reasoning. BICA 2010. (Submitted).Cassimatis, N.L., Trafton, J.G., Bugajska, M.D., & Schultz, A.C. (2004). Integrating cognition, perception and action through mental simulation in robots. Journal of Robotics and Autonomous Systems 49 (1-2): 13-23.Coward, L. A. (1990). Pattern Thinking. New York: Praeger.Murdock, J.W. and Goel, A.K. (2001). Meta-Case-Based Reasoning: Using Functional Models to Adapt Case-Based Agents. Proceedings of the 4th International Conference on Case-Based Reasoning (ICCBR'01). Vancouver, Canada, July 30 - August 2, 2001.Laird, J.E., Rosenbloom, P.S., & Newell, A. (1986). Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Boston: Kluwer. Laird, J.E., Newell, A., & Rosenbloom, P.S., (1987). SOAR: An architecture for general intelligence. Artificial Intelligence 33: 1-64.Th_risson, K. R. (1996). Communicative Humanoids: A Computational Model of Psycho-Social Dialogue Skills. Ph.D. Thesis, Media Laboratory, Massachusetts Institute of Technology. Th_risson, K. R. (1999). A Mind Model for Multimodal Communicative Creatures and Humanoids. International Journal of Applied Artificial Intelligence, 13(4-5): 449-486.
Other key referenceswww.sandia.gov/rohrerB. Chandrasekaran. Multimodal Cognitive Architecture: Making Perception More Central to Intelligent Behavior. Proceedings of the AAAI National Conference on Artificial Intelligence, 2006, pp. 1508-1512.Gobet, F. & Lane, P. (2005). The CHREST architecture of cognition: Listening to empirical data. In D. Davis (Ed.), Visions of mind: Architectures for cognition and affect. (pp. 204-224). Hershey, PA: Information Science Publishing. Kieras, D. EPIC Architecture Principles of Operation. ftp://www.eecs.umich.edu/people/kieras/EPICtutorial/EPICPrinOp.pdf Epstein, S. L. and S. Petrovic. 2010 Learning Expertise with Bounded Rationality and Self-awareness. In Autonomous Search. Y. Hamadi, E. M., F. Saubion, editors. Springer. www.numenta.comFranklin, S., & Ferkin, M. H. (2008). Using Broad Cognitive Models and Cognitive Robotics to Apply Computational Intelligence to Animal Cognition. In T. G. Smolinski, M. M. Milanova & A.-E. Hassanien (Eds.), Applications of Computational Intelligence in Biology: Current Trends and Open Problems (pp. 363-394): Springer-Verlag.Coward LA (2001) The Recommendation Architecture: lessons from the design of large scale electronic systems for cognitive science. Journal of Cognitive Systems Research 2:111-156; Coward LA (2005) A System Architecture Approach to the Brain: from Neurons to Conscious-ness. Nova Science Publishers, New York. SEE http://cs.anu.edu.au/~Andrew.Coward/References.htmlMurdock, J.W. and Goel, A.K. (2003). Localizing Planning with Functional Process Models. Proceedings of the Thirteenth International Conference on Automated Planning & Scheduling (ICAPS'03). Trento, Italy Ulam, P., Goel, A.K., Jones, J., and Murdock, J.W. (2005). Using Model-Based Reflection to Guide Reinforcement Learning. Proceedings of the IJCAI 2005 Workshop on Reasoning, Representation and Learning in Computer Games. Edinburgh, UKTh_risson, K. R. (1999). ?A Mind Model for Multimodal Communicative Creatures and Humanoids. International Journal of Applied Artificial Intelligence, 13(4-5):449-486. Th_risson, K. R. (2002). Machine Perception of Multimodal Natural Dialogue. P. McKevitt (Ed.), Language, vision & music . Amsterdam: John Benjamins. Th_risson, K. R. (1998). Real-Time Decision Making in Face to Face Communication. Second ACM International Conference on Autonomous Agents , Minneapolis, Minnesota, May 11-13, 16-23. Th_risson, K. R. (1997). Layered Modular Action Control for Communicative Humanoids. Computer Animation '97, Geneva, Switzerland, June 5-6, 134-143. Th_risson, K. R. (2002). Natural Turn-Taking Needs No Manual: A Computational Theory and Model, from Perception to Action. In B. Granstr_m (Ed.), Multimodality in Language and Speech Systems. Heidelberg: Springer-Verlag.
Most recent representative reference(s)Albus, J. S., and Barbera, A. J. (2005) "RCS: A cognitive architecture for intelligent multi-agent systems," Annual Reviews in Control, Vol. 29, Issue 1 87-99 Anderson, 2007http://cns.bu.edu/~stevewww.sandia.gov/rohrer B Rohrer, M Bernard, JD Morrow, F Rothganger, P Xavier, "Model-free learning and control in a mobile robot," 5th Intl Conf on Natural Computation, Tianjin, China, Aug 14-16, 2009. B Rohrer, JD Morrow, F Rothganger, PG Xavier, "Concepts from data," Brain-Inspired Cognitive Architectures Symposium, AAAI Fall Symposium Series, Washington DC, Nov 5-7, 2009. Arrabales, R. Ledezma, A. and Sanchis, A. "CERA-CRANIUM: A Test Bed for Machine Consciousness Research". International Workshop on Machine Consciousness. Hong Kong. June 2009.Gobet, F., Lane, P.C.R., Croker, S., Cheng, P. C-H., Jones, G., Oliver, I., and Pine, J. M. Chunking mechanisms in human learning. Trends in Cognitive Science, 5:236-243, 2001. Gobet, F., & Lane, P. C. R. (2010). The CHREST architecture of cognition: The role of perception in general intelligence. The Third Conference on Artificial General Intelligence. Lugano, Switzerland.http://www.cogsci.rpi.edu/~rsun/clarion.htmlGoertzel, Ben (2009). OpenCogPrime: A Cognitive Synergy Based Architecture for Embodied General Intelligence, Proceedings of ICCI-2009. See also http://opencog.org Goertzel, Ben et al. (2010). OpenCogBot: Achieving Generally Intelligent Virtual Agent Control and Humanoid Robotics via Cognitive Synergy, Proceedings of ICAI-10, BeijingEvertsz, R., Pedrotti, M., Busetta, P., Acar, H., & Ritter, F. E. (2009). Populating VBS2 with realistic virtual actors. In Proceedings of the 18th Conference on Behavior Representation in Modeling and Simulation. 1-8. 09-BRIMS-04.Epstein, S. L. and S. Petrovic. 2010 Learning a Mixture of Search Heuristics. Metareasoning in Thinking about thinking, Cox, M. T. and A. Raja, editors. MIT Press. Shapiro, S.C., & Bona, J.P. (in press). The GLAIR cognitive architecture. AAAI Technical Report FS-09-01. Menlo Park, CA: AAAI PressSamsonovich, A. V., Ascoli, G. A., De Jong, K. A., and Coletti, M. A. (2006). AAAI Technical Report WS-06-03, pp. 129?134. Menlo Park, CA: AAAI Press. George D, Hawkins J (2009) Towards a Mathematical Theory of Cortical Micro-circuits. PLoS Comput Biol 5(10)O'Reilly & Munakata (2000)Franklin, Stan. (2007). A foundational architecture for artificial general intelligence. In Advances in artificial general intelligence: Concepts, architectures and algorithms, Proceedings of the AGI workshop 2006, ed. Ben Goertzel and Pei Wang:36-54. Amsterdam: IOS Press.Rigid Flexibility: The Logic of Intelligence, Springer, 2006 and see http://sites.google.com/site/narswang/home Brom, C., Pe_kov_, K., Lukavsk_, J.: What does your actor remember? Towards characters with a full episodic memory. Proceedings of 4th International Conference on Virtual Storytelling, LNCS, SpringerCassimatis, N. L., Bignoli, P., Bugajska, M., Dugas, S., Kurup, U., Murugesan, A., & Bello, P. (in press). An Architecture for Adaptive Algorithmic Hybrids. IEEE Transactions on Systems, Man, and Cybernetics. Part B.Coward LA, Gedeon TO (2009) Implications of Resource Limitations for a Conscious Machine. Neurocomputing 72:767-788; Coward, L. A. (2010). The Hippocampal System as the Cortical Resource Manager: a model connecting psychology, anatomy and physiology. Advances in Experimental Medicine and Biology 657, 315 - 364.Murdock, J.W. and Goel, A.K. (2008). Meta-Case-Based Reasoning: Self-Improvement through Self-Understanding. Journal of Experimental & Theoretical Artificial Intelligence, 20(1):1-36Laird, 2008 (AGI)Ng-Thow-Hing, V., K. R. Th_risson, R. K. Sarvadevabhatla, J. Wormer and Thor List (2009). Cognitive Map Architecture: Facilitation of Human-Robot Interaction in Humanoid Robots. IEEE Robotics & Automation Magazine, March, 16(1):55-66. Jonsdottir, G. R.
Implementationhttp://www.isd.mel.nist.gov/projects/rcslib/ C++, Windows Real-time, VXworks, Neutral Messaging Language (NML), Mobility Open Architecture Simulation and Tools (MOAST) http://sourceforge.net/projects/moast/ Urban Search and Rescue Simulation (USARSim) http://sourceforge.net/projects/usarsim/Lisp, TCL/TkNonlinear neural networks (feedback, multiple spatial and temporal scales)MATLAB, PythonImplemented in the Soar framework.Net Framework, code written in C#. Runtime based on CCR (Concurrency and Coordination Runtime) and DSS (Decentralized Software Services), part of Robotics Developer Studio 2008 R2. http://www.conscious-robots.com/en/robotics-studio/2.html Lisp, JavaJavaThe OpenCogPrime system implements CogPrime within the open-source OpenCog AI framework, see http://opencog.org. The implementation is mostly C++ for Linux, some components in Java; also a Scheme shell is used for interacting with the system.overlay to JACK?Disciple was initially implemented in Lisp and is currently implemented in JavaOriginal version: Common Lisp; EPIC-X: C++Common Lisp, JavaCommon LispMatlab, PythonNuPIC development environment available for PC and Mac. It is available under a free research license and a paid commercial license."Emergent" open source software written largely in C++ (runs on many platforms)Extensible framework in Java.Open source, in Java and PrologIn Progress (C++, Java, Perl, Prolog)Java, Python; features ACT-R binding and Emergent bindingJavaSmalltalkCommon Lisp, using Loom as the underlying knowledge-engine; Java version in developmentC with interfaces to almost any language; JavaCommonLisp, C, C++, 8 networked computers, sensing hardware
Funding program, project and environment in which the architecture was applied (added by Jim Albus)Please see the long list: http://members.cox.net/bica2009/cogarch/4DRCS.pdfSandia National Labs, internal R&D fundingOffice of Naval ResearchNational Science FoundationDARPA IPTO BICA, virtual indoor/outdoor environmentsNumenta licesnses its software to various commercial partners.NSF Science of Design program; Self Adaptive Agents project; Turn-based strategy games
Working memory?yesNot explicitly definedYes. Recurrent shunting on-center off-surround network that obeys LTM Invariance Principle and Inhibition of Return rehearsal lawYes. Current percepts are folded together with a decayed version of recent percepts. This creates a short history of salient percepts that functions as working memory.Diagrammatic working memory addedyesYes, although we call it short-term memory. Auditory short-term memory and visuo-spatial short-term memory are implemented.separate structureyesThe active beliefs and intentionsThe framework of Disciple supports features and components that are common for many cognitive architectures, including working memory (reasoning trees), semantic memory (ontologies), episodic memory (reasoning examples), and procedural memory (rules). Communication is based on natural language patterns learned from the user.YesYesyes: includes mental states of the SelfMany Leabra working memory models have been published, mostly focusing on the role of the prefrontal cortex in working memory.Explicitly included as a workspace with significant internal structure including a current situational model with both real and virtual windows, ans a conscious contents queue, the active part of the memoryYesdeclarativeEach specialist implement its own.Yes. Frequency modulation placed on neuron spike outputs, with different modulation phase for different objects in working memory.No commitment to a specific theory of working memoryRelational graph structureFunctional Sketchboard, Content Blackboard, Motor Feedback Blackboard, Frames
Semantic memory?yesEncoded as chuncksYes, limited: Associations between chunksYes. Semantic information for a feature is obtained from the history of the agent's experience. no (to be implemented)Yes. Implemented by the network of chunks in long-term memory.yes. in both implicit and explicit forms (chunks/rules and NNs)yesencocoded as (weighted) beliefs in beliefsets, uses ACT-R's declarative memory equationsNot explicitlyIn many descriptivesYes, in SNePS yes: includes schemassemantic meaning can be encoded in sparse distributed representationsMany Leabra models of the learning and use of semantic knowledge, abstracted from the statistical regularities over many experiences, have been published. These include some language models. Implemented automatically as part of declarative memory via sparse distributed memorythe whole memory is semanticframesnoneEach specialist implement its own.Yes. Similarity circumstances (= cortical column receptive fields) that are often detected at the same time acquire ability to indirectly activate each other.Uses Powerloom as underlying knowledge engine; OWL based ontological representationRelational graph structuresFrames
Episodic memory?yesNot explicitly definedYes, limited: Builds on hippocampal spatial and temporal representationsYes. Individual transitions from experience to experience are an episode.no (to be implemented)No, although ?episodic? links are used in some simulations. yes. in both implicit and explicit forms (chunks and NNs)yesNot explicetly defined, but would be encocoded as beliefs in beliefsets.NoYes, in task history and summary measurementsYes, temporally-related beliefs in SNePS yes: includes frozen mental state assembliesMany Leabra models of episodic memory have been published, mostly focusing on the role of the hippocampus in episodic memory.Both declarative memory and transient episodic memory encoded via sparse distributed memorythe part of memory that contains temporal informationYes (Implicit)declarative or a spreading activation networkEach specialist implement its own.Yes. Similarity circumstances (= cortical column receptive fields) that change at the same time acquire ability to indirectly activate each other. An episodic memory is indirect activation of a group of columns that all changed at the same time at some point in the past. Because the hippocampal system manages the selection of cortical columns that will change in response to each sensory experience, information used for this management is applicable to construction of episodic memories Yes: Defined as traces through the procedural memory, belowEncoded as graph structures (snapshots of working memory)Functional Sketchboard, Content Blackboard, Motor Feedback Blackboard
Procedural memory?yesExplicitly definedYes. Multiple explicitly defined neural systems for learning, planning ,and control of actionYes. Common sequences of transitions are reinforced and are more likely to be executed in the future.Extends Soar's procedural memory to diagrammatic componentsyes. Some processors generate single or complex behaviors.Yes. Implemented by productions. yes. in both implicit and explicit forms (chunks/rules and NNs)yesPlans and intentions with activation levelsYes: production rulesYes, Advisors can be weighted by problem progress, and repeated sequences of actions can be learned and storedYes, in PML implemented in Lisp, could be compiled from KL yes: includes primitivesA fair number of Leabra models of automatized sequential action have been produced, with a smaller number specifically addressing issues of motor control. Most of these models explore the shaping of distributed patterns of synaptic strengths in posterior brain areas in order to produce appropriate action sequences in novel situations. Some work on motor skill automaticity has been done. A few models, integrating prefrontal and posterior areas, have focused on the application of explicitly provided rules.Schemas a la Drescherthe part of memory that is directly related to executable operationsYes (Implicit)rulesEach specialist implement its own.Yes. Recommendation weights associated in the basal ganglia with cortical column receptive field detections instantiate procedural memoryYes: Defined as tasks, which are functional elements, and methods, which are behavioral elementsRulesFrames, Action Modules, (limited implementation)
cognitive map?yesYes. Networks that learn entorhinal grid cell and hippocampal place field representations on lineNo explicit. The set of all transitions form a world model. This serves as an implicit cognitive map.Cognitive map emergesyes. Mission-specific processors build 2D maps.No.yesyesLeabra contains the mechanisms necessary to self-organize topographic representations. These have been used to model map-like encodings in the visual system. At this time, it is not clear that these mechanisms have been successfully applied to spatial representation schemes in the hippocampus. as part of the memorygraph based and BayesianThere is a spatial secialist that has this functionality.No. Requirement to conserve resources by using any one similarity circumstance (= receptive field) to support multiple behaviours precludes the existence of unambigous cognitive mapsNo commitment to a specific theory of cognitive mapsLimited body-centric spatial layout of selected objects
reward system?Yes, Value Judgment processes compute cost, benefit, riskYes. Model how amygdala, hypothalamus, and basal ganglia interact with sensory and prefrontal cortex to learn to direct attention and actions towards valued goals. Used to help explain data about classical and instrumental conditioning, mental disorders (autism, schizophrenia), and decision making under risk.Yes. Reward is specified in the definition of the task. It can be an arbitrary function of observed and unobserved states.yes. Status assessment mechanism in core layer.yes. in the form of a motivational subsystem and a meta-cognitive subsystem (MCS determines rewards based on the MS)Yes, Value Judgment processes compute cost, benefit, riskUses ACT-R memory equations, so memories and plans get strengthened.NoYes; Advisor weights are acquired during self-supervised learning.yesLeabra embraces a few alternative models of the reward-based learning systems dependent on the mesolimbic dopamine systems, including a neural implementation of temporal difference (TD) learning, and, more recently, the PVLV algorithm. Models have been published involving these mechanisms, as well as interactions between dopamine, the amygdala, and both lateral and orbital areas of prefrontal cortex.Feeling & emotion nodes in perceptual associative memoryexperience-based and context-sensitive evaluationYes. Some receptive field detections are associated with recommendations to increase or decrease recently used behavioural recommendation weightsFunctional descriptions of tasks allow agents to determine success or failure of those tasks by observing the state of the world. Success or failure can then be used to reinforce decisions made during execution.appraisal-based reward as well as user-defined internal/external rewardNo
Iconic memory (including Interface and Imagery)?Yes, image and map representationsPropositional (Based on chunks)Emerges from role of top-down attentive interactions in laminar models of how the visual cortex seesNo explicit. An implicit iconic memory emerges from the transision history between vision-based features. no (to be implemented)Yes.Intended to emerge implicitly via combination of sensory, declarative and simulative memory. Not implemented/tested yet.not particularly, would be application and IO system specific.Part of visual perceptual processorYes, in some descriptivesyes (imagery was not implemented)In Leabra, iconic memory can result from activation-based attractor dynamics or from small, sometimes transient, changes in synaptic strength, including mechanisms of synaptic depression. Imagery naturally arises from patterns of bidirectional excitation, allowing for top-down influences on sensory areas. Little work has been done, however, in evaluating Leabra models of these phenomena against biological data.Defined using various processed pixel matricesNot directly implemented. Extension possible.not sureThere is a spatial secialist that has some of this functionality.Yes. Maintained on the basis of indirect activation of cortical columns recently active at the same timeSpatial relationships are encoded using the underlying knowledge engine.Explicitly definedNo
Perceptual memory (if understood separately from iconic and working memory)?Yes, pixels are segmented and grouped into entities and eventsYes. Model development of laminar visual cortex and explain how both fast perceptual learning with attention and awareness, and slow perceptual learning without attention or awareness, can occur.Yes. All different sensory modalities and combinations of modalities are treated the same within BECCA.Yes. Diagrammatic elements of perceptual memnoryYes, but only as preprocessing buffers in CERA sensor services.Yes.Yes,but currently only for vision. Further modalities will be implemented later.gets input from the world as events. these events are processed by plans.Yesyes: input-output buffer (in psychology, very short-term perceptual memory and iconic memory are synonyms)Different aspects of perceptual memory can be supported by activation-based learning, small changes in synaptic strengths, frontally-mediated working memory processes, and rapid sparse coding in the hippocampus.Semantic net with activation passing. Nodes may have sensory data of various sorts attached.There is a spatial secialist that has some of this functionality.Yes. Based on indirect activation of cortical columns on the basis of recent simultaneous activityPerceptual representations at multiple levels (complexity) and timescales (see ?visual input? and ?auditory input? below).
Attention and consciousnessYes, can focus attention on regions of interest. Is aware of self in relation to the environment and other agents.Yes. Clarifies how boundary, surface, and prototype attention differ and work together to coordinate object and scene learning. Adaptive Resonance Theory predicts a link between processes of Consciousness, Learning, Expectation, Attention, Resonance, and Synchrony (CLEARS) and that All Conscious States Are Resonant States.Yes. The most salient feature at each time step is attended.Uses Soar's attention frameworkYes, attention is implemented as a bias signal induced from the core layer to the lower levels global workspaces. Other aspects of consciousness are also considered. No explicit representation of the self implemented so far.Attention plays an important role in the architecture, as it for example determines the next eye fixation and what will be learnt.Yes, can focus attention on regions and topics of interest. Is aware of self in relation to the environment and other agents.Represented by transient events, goal structure, and intentions/beliefs whose activation level is above the thresholdEmergent Phenomenon of working memoryYes, some Advisors attend to specific problem featuresDistribution of attention is described by a special attribute in instances of schemas. Consciousness can be identified with the content of the mental state "I-Now".We have modeled covert attentional mechanisms within HTMs although this is not in currently released product.In Leabra, attention largely follows a "biased competition" approach, with top-down activity modulating a process that involves lateral inhibition. Lateral inhibition is a core mechanism in Leabra, as is the bidirectional excitation needed for top-down modulation. Models of spatial attention have been published, including models that use both covert shifts in attention and eye movements in order to improve object recognition and localization. Published models of the role of prefrontal cortex in cognitive control generally involve an attention-like mechanism that allows frontally maintained rules to modulate posterior processing. Virtually no work has been done on "consciousness" in the Leabra framework, though there is some work currently being done on porting the Mathis and Mozer account of visual awareness into Leabra.Implemented a la Global workspace theory with global broadcasts recruiting possible actions in reponse to the current contents and, also, modulating the various forms of learning.Atttntion via expectation generation and matching.Attention.Yes. Cortical columns have receptive fields defined by groups of similar conditions that often occurred at the same time in past sensory experiences, and are activated if the receptive field occurs in current sensory inputs. Attention is selection of a subset of currently detected columns to be allowed to communicate their detections to other cortical areas. The selection is on the basis of recommendation strengths of active cortical columns, interpreted through the thalamus, and is implemented by placing a frequency modulation on the action potential sequences generated by the selected columns. Consciousness is a range of phenomena involving pseudosensory experiences. A column can also be indirectly activated if it was recently active, or often active in the past, or if it expanded its receptive field at the same time as a number of currently active columns. Indirect activations lead to conscious experiences. Implicit in the fact that methods constrain how and when subtasks may be executed (via a state-machine); each method can be in only one state at a time, corresponding to the attended-to portion of a reasoning task (and directly linked to the attended-to knowledge via the tasks requirements and effects).Within-utterance attention span. Situated spatial model of embodied self (but no semantic representation of self that could be reasoned over).
Visual input?Yes, color, stereoPropositional (based on chunks)Natural static and dynamic scenes, psychophysical displays. Used to develop emerging architecture of visual system from retina to prefrontal cortex, including how 3D boundaries and surface representations form, and how view-dependent and view-invariant object categories are learned under coordinated guidance of spatial and object attention.YesYes. Diagrams are the visual input.Yes, both real cam and synthetic images from the simulator.Visual input is currently coded as list structures or arrays.Currently handled via interfacing with external vision processing tools. Tighter interlinkage with a hierarchical NN vision system is a topic of current research.currently depends on the particular modelInteraction of visual motor and visual perceptual processorsPossible but not implementedperceptual structures in PMLsymbolicHTMs are inherently modality independent although we have applied them to vision tasks. We offer a Vision Framework for programmers and a Vision Toolkit requiring no programming skills.An advanced Leabra model of visual object recognition has been produced which receives photographic images as input.possible but not implementedYes. Both raw video and symbolic representations of raw visual inputs.both symbolic and subsymbolic (tailored for the purposes ov virtual worlds)Currently implemented by emulation of action potential outputs of populations of simulated sensory neuronsNot implemented, but would be handled by the underlying knowledge enginePropositional or relationalYes. Temporally and spatially accurate vector model of upper human body, including hands, fingers, one eye. Via body-tracking suit, gloves and eyetracker.
Auditory input?No, not yetPropositional (based on chunks)Natural sound streams. Used to develop emerging architecture of visual system for auditory streaming and speaker-invariant speech recognitionYesNono (to be implemented)Auditory input is currently coded as text input (segmented either as words, phonemes, or syllables).Currently the only auditory input we handle is speech, via an external speech-to-text engine. The architecture supports it in principle.Auditory and speech objects, spatialized auditory informationYes. FORRSooth is an extended version that conducts human-computer dialogues in real timeHas been done using off-the-shelf speech recognitiontextual commandsWe have worked with customers applying HTMs to auditory tasks.While a few exploratory Leabra models have taken low-level acoustic features as input, this modality has not yet been extensively explored.possible but not implementedNo.noneCurrently implemented by emulation of action potential outputs of populations of simulated sensory neuronsNot implemented, but would be handled by the underlying knowledge engineSupport for text-based communicationYes. Speech recognition (BBN Hark), custom real-time prosody tracker with H* and L* detection.
Special modalities?Yes, LADAR, GPS, odometry, inertialYes. SAR, LADAR, multispectral IR, night vision, etc.BECCA is modality agnostic. It can handle inputs originating from any sensory modality.SONAR, Laser Range Finder.Reads text from the Internet ;-)Accepts data from external databasesagents have used speech and navigationHTMs have been applied to vision, audition, network sensors, power systems, and other tasks.All senses are special, are they not?no built-in modalities, but allow plug-in sensors and actuatorsPossible to extend.No Multimodal integration and realtime multimodal communicative act interpretation
Reinforcemenent learningYes. Learns parameters for actuator backlash, braking, steering, and acceleration.Yes, for productions (linear discount version)Yes: CogEM and TELOS models of how amygdala and basal ganglia interact with orbitofrontal cortex etcBECCA is an RL algorithm and represents a novel approach to RL.Extends Soar's chunking for diagrammatic componentsno (could be implemented in specific processors)No.yesYes. Strengthening/weakening of plansAn expert interacts directly with a Disciple cognitive assistant, to teach it to solve problems in a way that is similar to how the expert would teach a less experienced collaborator. This process is based on mixed-initiative problem solving (where the expert solves the more creative parts of a problem and the agent solves the more routine ones), integrated learning and teaching (where the expert helps the agent to learn by providing examples, hints and explanations, and the agent helps the expert to teach it by asking relevant questions), and multistrategy learning (where the agent integrates complementary learning strategies, such as learning from examples, learning from explanations, and learning by analogy, to learn general concepts and rules).NoYesyes (a version of it)Leabra embraces a few alternative models of the reward-based learning systems dependent on the mesolimbic dopamine systems, including a neural implementation of temporal difference (TD) learning, and, more recently, the PVLV algorithm. Models have been published involving these mechanisms, as well as interactions between dopamine, the amygdala, and both lateral and orbital areas of prefrontal cortex.Yes, for perception, episodic and procedural memories via base-level activationNot decided.Yes. Increases in recommendation weights associated in the basal ganglia with cortical column receptive field detections, on the basis of rewards.Yes. Learns criteria for selecting alternative methods for accomplising tasks, and also alternative transitions within a methods state-transition machine.Yes, for operators (SARSA/Q-learning)?Yes, in a recent implementation (RadioShowHost)
Bayesian UpdateNot implemented, but could be.Yes, for memory retrievalNo, sort of: Includes some Bayes effects as emergent propertiesNono (could be implemented in specific processors)No.YesRetrieval of beliefs and access to current intentionsNoYesnoHTM hierarchies can be understood in a belief propagation/Bayesian framework.While Leabra does not include a mechanism for updating knowledge in a Bayes-optimal fashion based on singular experiences, it's error-driven learning mechanism does approximate maximum a posteriori outputs given sufficient iterated learning.NoYes.NoNot implemented, but could be.NoNo
Hebbian LearningNot implemented, but could beNoYes, sort of: Hebbian learning law is insufficient. Both Hebbian and anti-Hebbian properties are needed.Not formally, although the principle of associating co-occurring signals is used extensively.no (could be implemented in specific processors)No.yesYesNoYes, with respect to groupings of Tier-3 Advisors.yesYesAn associational learning rule, similar to traditional Hebbian learning, is one of the core learning mechanisms in Leabra.NoNo.the episodic memory features hebbian learningYes, but with an overlay management that determines whether Hebbian learning will occur at any point in timeNot implemented, but could beNoNo
Gradient Descent Methods (e.g., Backpropagation)Not implemented, but could beNoYes, but not BackpropagationNo. BECCA does not use artificial neural networks.no (could be implemented in specific processors)No.yesWe don't use this sort of algorithm explicitly, noNoNonoA biologically plausible error-correction learning mechanism, similar in performance to the generalized delta rule but dependent upon bidirectional excitation to communicate error information, is one of the core learning mechanisms in Leabra.NoNoNot implemented, but could beNoNo
Learning of new representationsYes. Learns maps and trajectories in new environmentsProduction compilation (forms new productions)Yes. Multiple kinds of self-organizationYesno (to be implemented)Yes. Chunks and templates (schemata) are automatically and autonomously created as a function of the interaction of the input and the previous state of knowledge.new chunks, new rules, new NN representationsYes, although this has been tested only in simple cases.NoYes, can learn new Advisors yesYesAll "active" representations in Leabra are, at their core, patterns of neural firing rates. These vectors of activity may be interpreted -- encoded and decoded -- in different ways, however. By analogy, all digital computer representations are strings of bits, but they may be interpreted as structures and pointers and the like. The way in which vectors of activity at "hidden" layers of neural units are interpreted is almost always a matter of learning in Leabra models. In this way, internal representations are always learned from experience.New representations (instrutionalist lerning) for perception, episodic and procedural memoriesCapable of representing newly learned knowledge.Yes. A new representation is a new subset of receptive field detections, with some slight changes to some receptive fieldsYes. Learns refinements of methods for existing tasks and can respond to a specification of some new task by adapting the methods for some similar task.Chunking (forms new rules); also mechanisms to create new episodes, new semantic memoriesNo
Main general paradigmsVisual and auditory information processingGoal-directed behaviorGlobal Workspace Theory. Multimodal sensory binding. Learning (e.g. implicit learning, verbal learning); acquisition of first language (syntax, vocabulary); expertise; memory; some problem solving; concept formationControl of virtual-world agents. Natural language processing. We are now starting to work with humanoid robots but that's early-stage.BDIDisciple agents have been developed for a wide variety of domains, including manufacturing [74], education [75], course of action critiquing [76], center of gravity determination [77,78], and intelligence analysis [79]. The most recent Disciple agents incorporate a significant amount of generic knowledge from the Science of Evidence, allowing them to teach and help their users in discovering and evaluating evidence and hypotheses, through the development of Wigmorean probabilistic inference networks that link evidence to hypotheses in argumentation structures that establish the relevance, believability and inferential force of evidence [80]. Disciple agents are used in courses at various institutions, including US Army War College, Joint Forces Staff College, Air War College, and George Mason University [numbers refer to the above citations].Human performance, multi-tasking, PRP Procedure, air traffic controlConstraint solving; game playing; robot navigation; spoken dialoguereasoning, belief changevoluntary perception, cognition and actionLearning, learning, learning. Take biological constraints seriously. Build more controlled processes on a foundation of more automatic processes.Global Workspace Theoryreasoning with insufficient knowledge and resourcesExpectation generation and matching via learing and reasoning on stored knowledge and sensory inputs.3D virtual worldsReasoningAll cognitive processes are implemented through sequences of receptive field activations, including both direct detections and indirect activations. At each point in the sequence the behaviour with the predominant recommendation weight across the currently activated receptive field population is performed. This behaviour may be to focus attention on a particular subset of current sensory inputs or to implement a particular type of indirect activation (prolong current activity, or indirectly activate on the basis of recent simultaneous activity, past frequent simultaneous activity, or past simultaneous receptive field change). Recommendation weights are acquired through rewards that result in effective sequences for cognitive processing. Frequently used sequences are recorded in the cerebellum for rapid and accurate implementation.Reflection. Adaptation in response to new functional requirements.Integrated behavior-based and classical AI; blackboards; distributed implementation
other general paradigmsPerception, abstractionVerbal working memory, visual working memoryreasoning without a goalself-regulated learning (at the stage of design)Tripartite brain organization.Modle findingPlanning, Reinforcement learningModular architecture ("schema-style") allows easy expansion of common and custom features and principles. Solves to some extent the scaling problem.
Problem SolvingUses rules and/or search methods for solving problemsYesYes, depending on what is meantYesYes, problem solving is the main application.yes, implicit.Yes. At moment, CHREST solves problems mostly by pattern recognition.yesYesYesNoYesyesIn the traditional AI meaning of "problem solving", involving the generation of a sequential plan to meet a novel goal, little work has been done in Leabra. Some Leabra models of sequential action can generalize when performing in a novel situation, but none of these models have addressed the traditional AI planning problem.YesYesA simple example is fitting together two objects. First step is activating receptive fields often active in the past shortly after fields directly activated by one object were active. Because objects have often been seen in the past in several different orientations, this indirect activation is effectively a "mental rotation". Receptive fields combining information from the indirect activation derived from one object and the direct activation from the other object recommend movements to fit the objects together. A bias is placed upon acceptance of such behaviours by taking on the task. YesYesNo
Decision MakingMakes decision based on Value Judgment calculationsYes?YesYesyesYes.yesYesYesNoYesyesMuch work has been done on Leabra modeling of human decision making in cases of varying reward and probabilistic effects of actions, focusing on the roles of the dopamine system, the norepinepherine system, the amygdala, and orbito-frontal cortex.YesUses Inference (both statistical and logical).There can be extensive indirect activation steps, with (slight) changes to receptive fields at each step. Eventually, one behaviour has a predominant recommendation strength in the basal ganglia, and this behaviour is the decision.Yes, specifically in selecting methods to perform a task and selecting transitions within a methodYesYes, using hierarchical decision modules as well as traditional planning methods
AnalogyNot implemented Yes in rule discovery applicationsYesnot implementedNo.Yes NoYes, via pattern matchingyesSome preliminary work has been done on using dense distributed representations in Leabra to perform analogical mapping.NoSomewhatYes, using the functional specification (requirements and effects) of tasks.LimitedNo
Language ProcessingNot implemented Yes?Yesnot implementedOnly acquisition of language.Comprehension and generation fully implemented. Dialogue is something we're actively working on. LimitedYesYesMany Leabra language models have been produced, focusing on both word level and sentence level effects.Beginning stageYes.YesYesNot implementedYesYes.
Working Memory Tasks?YesYesNo.yesUnclear exactly what this means. The system does many tasks involving working memory.YesVisual and Verbal WM TasksUnclear exactly what this meansLeabra models of the prefrontal cortex have explored a variety of working memory phenomena.YesYesYesYes.
perceptual illusionsNo.Noyes: modeled perceived flipping of the Necker cubeNoYesNo.
implicit memory tasksYes.yesNoIn principleYes. Depend on indirect receptive field activations on the basis of recent simultaneous activity
metacognitive tasksNo.yesNoYesyes: modeled perceived flipping of the Necker cubeNOYesNo.
social psychology tasksNo.yesNoNo
personality psychology tasksNo.yesNoPersonality and emotion can be modeled through Advisorsyes: modeled perceived flipping of the Necker cubeNoNo.
motivational dynamicsNo.yesNoNo
Stroop?Yes (multiple models)?YesnoNo.We aren't doing modeling of human cognition and so we haven't tried the system on standard ?cognitive modeling test problems. So no.noYes.Not yetNo?No.
Task SwitchingYesYes (multiple models)YesYesnot yetNo.yesSee aboveYesyesYes.Not yetYes (necessary feature for nexting)YesNot implemented?Yes.
Tower of Hanoi/LondonYesYesNoNo.yesSee aboveYesnoI think there might have been some preliminary work on Tower of London, but I'm not sure.Not yetYesYesNo.
PRP?YesNo.See aboveyesNot that I know of.Not yetNo?
Dual TaskYesYesNo.yesSee aboveyesNot that I know of.Not yetNo.YesNot implementedYesNo.
N-Back?YesYesNo.See aboveyesYes, but there is still work to be done, here.Not yetNo?
visual perception with comprehensionYesYesnot yetYes.We have done this in the virtual world. Currently research aims at doing it for humanoid robots alsoyesyesYesA powerful object recognition model has been constructed.Not yetYes.YesNot implementedYes (see ?special modalities? above)
spatial exploration, learning and navigationYesYesYesYesyesTo some extent.yesSee abovenoYesyesPreliminary work only.Not yetYesYesimplemented but not compared to human behaviorNo.
object/feature search in an environmentYes. Search for targets in regions of interest. YesYesnot imeplementedYes.See abovenoyesObject localization naturally arises in the object recognition model.Not yetyes.Yes. Yes.
learning from instructionsYes. Learn from subject matter experts.No, unless you mean supervised learningNot imeplementedNo.yesYes, in the virtual world contextnoYesYesquasi-implementedSome preliminary work on instruction following, particularly in the domain of classification instructions, has been done in Leabra.Not yetyes.Yes. Yes, from new task specificationsimplemented but not compared to human behaviorNo.
pretend-playNoYes: Outline of architecture for teacher-child imitationno No.See abovenoquali-implementedNot that I know of.Not yetsomeNo.
Uses only local computations?No. Uses information from battlefield information network, apriori maps, etc.Yes. All operations defined by local operations in neural networks.NoYes.No. Mix of local and globalYesyesYes, HTM is a biological model.Yes, throughout the architectureYes, in the existing implementation?
Unsupervised learning?Yes. Updates control system parameters in real-timeYes. Can categorize objects and events or alter spatial maps and sensory-motor gains without supervision.YesyesYes.Yes. noYesyesYes, uses time as primary learning method.Perceptual, episodic and procedural, each in both instructionalist and selectionist modesyesYes. Adapts in response to failures through situated action + reinforcement learning, or through generative planning + abstraction.No.
Supervised learning?Yes. Learns from subject matter experts.Yes. Can learn from predictive mismatches with environmental constraints, or explicit teaching signals, when they are available. Nonot implementedYes.Yes. noYesyesYes, networks can be supervised at top node.NoyesYes, from new task specificationsNo.
Arbitrary mixtures of unsupervised and supervised learning?YesYes. E.g., the ARTMAP family of models.Nonot implementedYes.YesnoYesyesNoyesYes. Can develop a new method from a task specification and then later adapt that method based on experience.No.
Can it learn in real time?YesYes. Both ART (match learning) and Vector Associative Map (VAM; mismatch learning) models use real-time local learning laws.YesyesYes.YesnoYesyesHTMs can learn on-line, meaning learning while inferring. Real time depends on size of network and nature of problem.YesSometimes.YesNo.
Can it do fast stable learning; i.e., adaptive weights converge on each trial without forcing catastrophic forgetting?Yes. Uses CMAC algorithm that learns fast from error correction. When no errors, learning stops.Yes, theorems prove that ART can categorize events in a single learning trial without experiencing catastrophic forgetting in dense non-stationary environments. Mismatch learning cannot do this, but this is adaptive in learning spatial and motor data about changing bodies.YesYes.Yes noYesin short, yes. Learning in GMU-BICA consists in storage of mental states and in creation of new schemas, without forgetting. Learning "weights" applies to the neuromorphic cognitive map. In most cases, learning occurs in one shot.YesYesNo.
Can it function autonomously?Yes. Operates machines and drives vehicles autonomously.Yes. ART models can continue to learn stably about non-stationary environments while performing in them.YesyesYes.Yes. Yes. Key motivation for selecting BDIyesYesyes.YesYesYesYes; however, no actual implementations have yet pushed the architecture on this issue.
Is it general-purpose in its modality; i.e., is it brittle?It is general purpose and robust in real world environments.ART can classify complex non-stationary data streams. The FACADE 3D vision model clarifies how multiple types of visual data (e.g., edge, texture, shading, stereo) are processed by laminar cortical circuits.BECCA is designed to be general purpose and robust to large and small changes in tasks.yesIt is general-purpose and not brittle.It is general purpose and robust in real world environments.noGeneral purposegeneral-purpose.YesGeneral purposeGeneral purposeThe architecture to some extent addresses the brittleness problem, but not to full autonomy.
Can it learn from artbitrarily large databases; i.e., not toy problems?Yes. All applications are real-world and real-time.Yes. Theorems about ART algorithms show that they can do fast learning and self-stabilizing memory in arbitrarily large non-stationary data bases. ART is therefore used in many large-scale applications http://techlab.bu.edu/Plannednot testedYes. Simulations on the acquisition of language have used corpora larger than 350k utterances. Simulations with chess have used databases with more than 10k positions Discrimination networks with up to 300k chunks have been created.Yes. noYesin principle, yes.Yes???Yes, up to the limitations of the underlying knowledge engineNo.
Can it learn about non-stationary databases; i.e., environmental rules change upredictably?Yes. Battlefield environments change unpredictably.Yes. See above.YesYes.Yes. noUntested but we believe soin principle, yes.It can be implemented for on-line learning.???YesNo.
Can it pay attention to valued goals? Yes. ART derives its memory stability from matching bottom-up data with learned top-down expectations that pay attention to expected data. ART-CogEM models use cognitive-emotional resonances to focus attention on valued goals.YesyesYes, through attention mechanisms.YesyesYesyes.YesYes. Goals are encoded as the intended effects of tasks.Yes.
Can it flexibly switch attention between unexpected challenges and valued goals?Yes. Makes decisions about what is most important based on rules of engagement and situational awareness.Yes. Top-down attentive mismatches drive attention reset, shifts, and memory search. Cognitive-emotional and attentional shroud mechanisms modulate attention shifts.YesyesYes.Yes. Yes.yesYesyes: by switching roles of mental states.YesNot implementedYes.
Can reinforcement learning and motivation modulate perceptual and cognitive decision-making?Yes. Cognitive-emotional and perceptual-cognitive resonances interact together for this purpose.YesyesNo.YesYes, cognitivenoYesyes.YesYesYes.
Can it adaptively fuse information from multiple types of sensors and modalities?ART categorization discovers multi-modal feature and hierarchical rule combinations that lead to predictive success.YesyesYes.Yes, in principle, but this has not been tested except on vision and language and speech input...noYesyes: by using / creating appropriate schemas.YesIn principle, but not implementedIn principle, but not implementedYes, although it depends on the particular use of the architectural features.
Most statements included here were written as summaries of the BICA-2009 CogArch panel presentations. Their short versions will be included in the AI Magazine symposium report.Biologically-relevant cognitive architectures should clarify how individuals adapt autonomously in real time to a changing world filled with unexpected events; should explain and predict how several different types of learning (recognition, reinforcement, adaptive timing, spatial, motor) interact to this end; should use a small number of equations in a larger number of modules, or microassemblies, to form modal architectures (vision, audition, cognition, ...) that control the different modalities of intelligence; should reflect the global organization of the brain into parallel processing streams that compute computationally complementary properties within and between these modal architectures; and should exploit the fact that all parts of the neocortex, which supports the highest levels of intelligence in all modalities, are variations of a shared laminar circuit design and thus can communicate with one another in a computationally self-consistent way. BECCA was designed to solve the problem of natural-world interaction. The research goal is to place BECCA into a system with unknown inputs and outputs and have it learn to successfully achieve its goals in an arbitrary environment. The current state-of-the-art in solving this problem is the human brain. As a result, BECCA's design and development is based heavily on insights drawn from neuroscience and experimental psychology. The development strategy emphasizes physical embodiment in robots.Chandrasekaran called attention to the lack of support in the current family of cognitive architectures for perceptual imagination, and cited his group's DRS system that has been used to help Soar and Act-R engage in diagrammatic imagination for problem solving.at the panel, I pointed out that the list of tasks needs to be greatly expanded, to include, for example, implicit learning tasks, meta-cognitive tasks, social psychology tasks, personality psychology tasks, motivational dynamics, and so on, all of which have been simulated using the CLARION cognitive architecture.The architecture described in this column was pioneered in the proprietary Novamente Cognition Engine, and is now being pursued in the open-source OpenCogPrime system, build with in the OpenCog framework... In my presentation in the BICA-2009 CogArch panel, I discussed the AGI Roadmap Initiative (see http://agi-roadmap.org) and also the need for a glossary of AGI terms to help us compare cognitive architectures.FORR (FOr the Right Reasons) is highly modular. It includes a declarative memory for facts and a procedural memory represented as a hierarchy of decision-making rationales that propose and rate alternative actions. FORR matches perceptions and facts to heuristics, and processes action preferences through its hierarchical structure, along with its heuristics? real-valued weights. Execution affects the environment or changes declarative memory. Learning in FORR creates new facts and new heuristics, adjusts the weights, and restructures the hierarchy based on facts and on metaheuristics for accuracy, utility, risk, and speed. Hierarchical Temporal Memory (HTM) is a model of neocortex and thalamus. It is highly constrained and guided by anatomy and physiology at the levels of cortical regions, cellular layers, cellular connectivity, local inhibitory neurons, and non-linear integration of synapses along dendrites. HTMs build hierarchical models of the spatial and temporal statistics in sensory data. The models can be used for inference and prediction. HTMs have been commercially applied to numerous commercial problems. Numenta aims to be a catalys t for commercial applications of neocortical models of compution and inference. It publishes its algorithms and code for research and commercial deployment.Though NARS can be considered as a "cognitive architecture" in a broad sense, it is very different from the other systems. Theoretically, NARS is a normative theory and model of intelligence and cognition as "adaptation with insufficient knowledge and resources", rather than a direct simulation of human cognitive behaviors, capabilities, or functions; technically, NARS uses a unified reasoning mechanism on a unified memory for learning, problem-solving, etc., rather than integrates different techniques in an architecture. Therefore, accurately speaking it is not after the same goal as many other cognitive architectures, though still related to them in various aspects.Nick Cassimatis argued that repositories should focus on standardizing task environments and problems within them rather than details about cognitive architectures themselves.Theoretical arguments indicate that any system which must learn to perfom a large number of different behaviors will be constrained into this recommendation architecture form by a combination of practical requirements including the need to limit information handling resources, the need to learn without interference with past learning, the need to recover from component failures and damage, and the need to construct the system efficiently. The actual systems built in the Ymir architecture, notably Gandalf, the Cognitive Map for Asimo, and the SuperRadioHost systems, have shown the Ymir framework to be quite flexible and extensible. Ymir was not initially built to solve general-purpose intelligence, and it is become clear that none of the Constructionist methodologies (i.e. most efforts to date relying on ?first order? manually-constructed software) will be able to handle the multitude of topics related to general intelligence, such as global attention, global learning, flexible task learning and integration, etc. For this we are working on new methodologies relying on Constructivist approaches, which emply self-organization and automatic architectural growth (see my keynote paper From Constructionist to Constructivist A.I. (2009), in AAAI Fall Symposium Series - Biologically Inspired Cognitive Architectures, Washington D.C., Nov. 5-7,175-183. AAAI Tech Report FS-09-01, AAAI press, Menlo Park, CA.).