A Standard Model of the MindJune 7, 201737th Soar Workshop.ppt

《A Standard Model of the MindJune 7, 201737th Soar Workshop.ppt》由会员分享，可在线阅读，更多相关《A Standard Model of the MindJune 7, 201737th Soar Workshop.ppt（19页珍藏版）》请在麦多课文档分享上搜索。

1、A Standard Model of the Mind June 7, 2017 37th Soar Workshop,John E. Laird, University of Michigan Christian Lebiere, Carnegie Mellon University Paul S. Rosenbloom, University of Southern CaliforniaSlides originally created by Paul S. Rosenbloom Based on Laird, J. E., Lebiere, C. & Rosenbloom, P. S.

2、 (2017). A Standard Model of the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics. AI Magazine. In press.,Functional entities that can think And thus support intelligent behavior Humans, and many animals possess minds Impleme

3、nted by brains In AI, minds assumed to be computational entities Cognitive systems Implementable via a range of physical devices,Minds,Artificial Intelligence (& AGI) Concerns building artificial minds Cares most for how such minds can be built Cognitive Science Concerns modeling natural minds Cares

4、 most for understanding human cognitive processes Neuroscience Concerns structure and function of brains Cares most for how minds arise from brains Robotics Concerns building and controlling artificial bodies Cares most for how minds control such bodies,Approaches to Understanding Minds,Deep scienti

5、fic question whether will converge on a single understanding of mind Must at least happen for cognitive science and neuroscience Naturally inspired AI/robotics may also fit if class slightly abstracted So also may other work that is similar for functional reasons Call this slightly abstracted class

6、human-like minds More the bounded rationality found in humans than AI optimality Broader than naturally inspired minds Narrower than human-level intelligence,Towards a Common Understanding,Integrative model Combines electromagnetic, weak and strong interactions Classifies all of the known elementary

7、 particles Community consensus developed over decades Assumed to be internally consistent but may have gaps Yields critical functionality for field Cumulative reference point for progress Guides efforts to extend and break it,A Standard Model of Particle Physics,AN ANALOGY,Focused on human-like mind

8、s Grew out of a summary presentation at the 2013 AAAI Fall Symposium on Integrated Cognition Yielded surprising consensus of those in attendance Organizing committee for Symposium drawn from cognitive science, cognitively and biologically inspired AI, AGI, and robotics Suggested a consensus has begu

9、n to emerge Followed up with an article for AI Magazine Extend consensus via a dialectic among ACT-R, Soar & Sigma Laird, J. E., Lebiere, C. & Rosenbloom, P. S. (2017). A Standard Model of the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscien

10、ce, and Robotics. AI Magazine. In press.,A Standard Model of The Mind (SMM),ACT-R, Soar & Sigma,ACT-R Parallel, asynchronous modules Central procedural memory Only one result/module/cycle Includes mapping to brain regions,Soar Parallel, asynchronous modules Central working memory Bipartite declarati

11、ve memory Visuospatial module,Sigma Blends CAs with graphical models Less modular architecturally Differing functionalities fromspecialization and combination,What needs to be in a human-like cognitive architecture Providing an abstract community consensus But not itself a cognitive architecture Yie

12、lds a range of potential benefits Coherent baseline to facilitate shared cumulative progress Focus efforts to extend and break the consensus Framework around which evaluation data can be organized Interlingua for describing and comparing architectural approaches Guidance in Extending research on ind

13、ividual components Interpreting experiments and suggesting new ones Constructing intelligent applications,A Standard Model of The Mind (cont.),Some Precedents*,Model Human Processor (Card, Moran & Newell, 1983),Scale Counts in Cognition (Newell, 1990),Generic Architecture for Human-Like Cognition (G

14、oertzel, Pennachin & Geisweiller, 2014)*,Very abstract Parameters for HCI,Span mind Deliberate Act up?,Amalgamated completeness CogPrime, CogAff, LIDA, Psi, 4D-RCS, HTM/DeSTIN rather than consensus,Simple reaction time in humans (roughly) Where cognitive architectures most commonly defined Although

15、may go a bit below Where standard model is defined Provides a physical symbol system “ has the necessary and sufficient means for general intelligent action” (Newell & Simon, 1976),Deliberate Act Level (100 ms),Key aspect of symbols for SMM is composability Symbols are those things that can be combi

16、ned into relations And support further combinations into networks of relations Mirrors binding problem in cognitive neuroscience Doesnt assume symbols must be uninterpreted labels Can be this and/or patterns over vectors of distributed elements Doesnt assume sufficiency at deliberate act level Logic

17、ally sufficient overall, but Evidence shows non-symbolic needed at deliberate act level To represent numeric and spatial information To provide quantitative metadata over symbol structures Frequency, recency, co-occurrence, similarity, utility, activation,Standard Model and Symbol Systems,Our unders

18、tanding of consensus from AAAI symposium fleshed out based on dialectic among three architectures Finding out what each actually meant and most believed A set of communicating modules Each may be decomposed further But not just undifferentiated stuff Global comms. through WM Local PerceptionMotor co

19、mms. Procedural can access all of WM Others access only own WM buffers At this level of detail, not too novel Although bipartite LTM is relatively new as a consensus,Standard Model,Structure and Processing Memory and Content Learning Perception and Motor,Standard Model (cont.),Processing yields boun

20、ded rationality, not optimality Processing based on a few task-independent modules There is significant parallelism in architectural processing Processing is parallel across modules Processing is parallel within modules A cognitive cycle that runs at 50 ms per cycle in humans drives behavior via seq

21、uential action selection Complex behavior arises from a sequence of independent cognitive cycles that operate in their local context,A. Structure and Processing,Declarative and procedural LTMs contain symbol* structures and associated quantitative metadata Global communication is provided by a short

22、-term WM Global control is provided by procedural LTM Composed of rule-like conditions and actions Exerts control by altering contents of WM Factual knowledge is provided by declarative LTM,B. Memory and Content,All forms of LTM content are learnable Learning occurs online and incrementally, as a si

23、de effect of performance and is often based on an inversion of the flow of information from performance Procedural learning involves at least reinforcement learning and procedural composition Reinforcement learning yields weights over action selection Procedural composition yields behavioral automat

24、ization Declarative learning involves the acquisition of facts and tuning of their metadata More complex forms of learning involve combinations of the fixed set of simpler forms of learning,C. Learning,Perception yields symbol* structures with associated metadata in specific WM buffers There can be

25、many different such perception modules Perceptual learning acquires new patterns & tunes existing ones Attentional bottleneck constrains information available in WM Perception can be influenced by top-down information from WM Motor control converts symbol* structures in its buffers into external act

26、ions There can be multiple such motor modules Motor learning acquires new action patterns & tunes existing ones,D. Perception and Motor,ACT-R & Soar disagreed in a number of ways decades ago Although disagreed about different things Now in essentially complete agreement on this consensus Although no

27、t all is completely implemented Sigma differs further in placing some functionality above architecture And less completely implemented at this point,Analysis of Three Architectures,Structure and Processing Memory and Content Learning Perception and Motor,Standard model reflects a very real initial c

28、onsensus Serial and parallel processing Symbolic and subsymbolic representations Procedural and declarative memories Pervasive learning But it remains incomplete in a number of ways Metacognition, emotion, mental imagery, distinction between declarative memories, social cognition, etc. And may still

29、 be wrong in various ways Hope is that it yields sound beginning to build upon via more Consensus across architectures, researchers, data and disciplines Work on challenges, corrections and extensions 2017 AAAI Fall Symposium on SMM To “begin the process of broader community involvement”,Conclusion,