A Neurally Plausible model of Reasoning.ppt

《A Neurally Plausible model of Reasoning.ppt》由会员分享，可在线阅读，更多相关《A Neurally Plausible model of Reasoning.ppt（59页珍藏版）》请在麦多课文档分享上搜索。

1、A Neurally Plausible model of Reasoning, Lokendra Shastri ICSI, Berkeley,Lokendra Shastri International Computer Science Institute Berkeley, CA 94704,Five levels of Neural Theory of Language,Cognition and Language,Computation,Structured Connectionism,Computational Neurobiology,Biology,SHRUTI,abstrac

2、tion,“John fell in the hallway. Tom had cleaned it. He got hurt.”,Tom had cleaned the hallway.The hallway floor was wet.John slipped and fell on the wet floor.John got hurt as a result of the fall.,such inferences establish referential and causal coherence.,Reflexive Reasoning,UbiquitousAutomatic, e

3、ffortlessExtremely fast - almost a reflex response of our cognitive apparatus,Reflexive Reasoning,Not all reasoning is reflexive Contrast with reflective reasoningdeliberateinvolves explicit consideration of alternativesrequire props (paper and pencil)e.g., solving logic puzzles differential equatio

4、ns,How fast is reflexive reasoning?,We understand language at the rate of 150-400 words per minuteReflexive inferences required for establishing inferential and causal coherence are drawn within a few hundred millisecond, Lokendra Shastri ICSI, Berkeley,How can a system of slow and simple neuron-lik

5、e elements,encode a large body of semantic and episodic knowledge and yet perform a wide range of inferences within a few hundred milliseconds?, Lokendra Shastri ICSI, Berkeley,Characterization of reflexive reasoning?,What can and cannot be inferred via reflexive processes?, Lokendra Shastri ICSI, B

6、erkeley,Shruti,http:/www.icsi.berkeley.edu/shastri/shrutiLokendra Shastri V. Ajjanagadde (Penn, ex-graduate student) Carter Wendelken (UCB, ex-graduate student) D. Mani (Penn, ex-graduate student) D.J. Grannes (UCB, ex-graduate student) Jerry Hobbs, USC/ISI (abductive reasoning) Marvin Cohen, CTI (m

7、etacognition; belief and utility) Bryan Thompson, CTI (metacognition; belief and utility),Lokendra Shastri ICSI, Berkeley,Reflexive Reasoning representational and processing issues,Activation-based (dynamic) representation of events and situations (relational instances),Dynamic representation of rel

8、ational instances,“John gave Mary a book”,giver: John recipient: Mary given-object: a-book,*,Reflexive Reasoning,Expressing dynamic bindings Systematically propagating dynamic bindings Computing coherent explanations and predictions evidence combination dynamic instantiation and unification of entit

9、ies,Requires compatible neural mechanisms for:,All of the above must happen rapidly,Learning,one-shot learning of events and situations (episodic memory) gradual/incremental learning of concepts, relations, schemas, and causal structures,Relation focal-cluster,+ - ? fall-pat fall-loc,FALL,Entity, ca

10、tegory and relation focal-clusters,+ - ? fall-pat fall-loc,FALL,Entity, category and relation focal-clusters,+ - ? fall-pat fall-loc,FALL,Functional nodes in a focal-cluster collector (+/-), enabler (?), and role nodes may be situated in different brain region,Focal-cluster of a relational schema,FA

11、LL,focal-clusters of motor schemas associated with fall,focal-clusters of lexical know- ledge associated with fall,focal-clusters of perceptual schemas and sensory representations associated with fall,focal-clusters of other relational schemas causally related to fall,episodic memories of fall event

12、s,Focal-clusters,Nodes in the fall focal-cluster become active whenperceiving a fall event remembering a fall event understanding a sentence about a fall event experiencing a fall event,A focal-cluster is like a “supra-mirror” cluster,Focal-cluster of an entity,John,+ ?,focal-clusters of motor schem

13、as associated with John,focal-clusters of lexical know- ledge associated with John,focal-clusters of perceptual schemas and sensory representations associated with John,focal-clusters of other entities and categories semantically related to John,episodic memories where John is one of the role-filler

14、s,+ - ? fall-pat fall-loc,Fall,+ ?,+ ?,Hallway,John,“John fell in the hallway”,+ - ? fall-pat fall-loc,Fall,+ ?,+ ?,Hallway,John,“John fell in the hallway”,+ - ? fall-pat fall-loc,Fall,+:Fall,+:John,fall-pat,fall-loc,+:Hallway,“John fell in the hallway”,Encoding “slip = fall” in Shruti,SLIP,FALL,+ -

15、 ? fall-pat fall-loc,Such rules are learned gradually via observations, by being told ,“John slipped in the hallway”,Slip,+ - ? slip-pat slip-loc,Fall,+ - ? fall-pat fall-loc,mediator,r2,r1,?,+, “John fell in the hallway”,A Metaphor for Reasoning,An episode of reflexive reasoning is a transient prop

16、agation of rhythmic activity Each entity involved in this reasoning episode is a phase in this rhythmic activity Bindings are synchronous firings of cell clusters Rules are interconnections between cell-clusters that support propagation of synchronous activity,Focal-clusters with intra-cluster links

17、,John,+ ?,+ - ? fall-pat fall-loc,FALL,+e +v ?v ?e,Person,Shruti always seeks explanations,Encoding “slip = fall” in Shruti,SLIP,FALL,+ - ? fall-pat fall-loc,+ ? r1 r2,mediator,Linking focal-clusters of types and entities,John,+e +v ?v ?e,+e +v ?v ?e,+e +v ?v ?e,Hallway,Location,Man,Person,Focal-clu

18、sters and context-sensitive priors (T-facts),+ ?,+ - ? fall-pat fall-loc,+e +v ?v ?e,*,context-sensitive priors,*,* cortical circuits,entities and types,entities and types,John,FALL,Person,Focal-clusters and episodic memories (E-facts),+ ?,+e +v ?v ?e,episodic memories,e-memories,e-memories,from rol

19、e-fillers,to role-fillers,*,*,*,* hippocampal circuits,John,FALL,Person,Explaining away in Shruti,FALL,+ ?,Other features of Shruti,Mutual inhibition between collectors of incompatible entities Merging of phases - unification Instantiation of new entities Structured priming,Unification in Shruti : m

20、erging of phases,The activity in focal-clusters of two entity or relational instances will synchronize if there is evidence that the two instances are the sameR1: Is there an entity A of type T filling role r in situation P? (Did a man fall in the hallway?) R2: Entity B of type T is filling role r i

21、n situation P.(John fell in the hallway.)In such a situation, the firing of A and B will synchronize. Consequently, A and B will unify, and so will the relational instances involving A and B.,Entity instantiation in Shruti,If Shruti encodes the rule-like knowledge:x:Agent y:Location fall(x,y) = hurt

22、(x)it automatically posits the existence of a location where John fell in response to the dynamic instantiation of hurt(x),Encoding “fall = hurt” in Shruti,FALL,HURT,+,-,?,hurt-pat, Lokendra Shastri ICSI, Berkeley,The activation trace of +:slip and +:trip,“John fell in the hallway. Tom had cleaned i

23、t. He got hurt.”,Lokendra Shastri ICSI, Berkeley,A Metaphor for Reasoning,An episode of reflexive reasoning is a transient propagation of rhythmic activity Each entity involved in this reasoning episode is a phase in this rhythmic activity Bindings are synchronous firings of cell clusters Rules are

24、interconnections between cell-clusters that support context-sensitive propagation of activity Unification corresponds to merging of phases A stable inference (explanation/answer) corresponds to reverberatory activity around closed loops,Support for Shruti,Neurophysiological evidence: transient synch

25、ro-nization of cell firing might encode dynamic bindings Makes plausible predictions about working memory limitations Speed of inference satisfies performance requirements of language understanding Representational assumptions are compatible with a biologically realistic model of episodic memory,Neu

26、rophysiological evidence for synchrony,Synchronous activity found in anesthetized cat as well as in anesthetized and awake monkey. Spatially distributed cells exhibit synchronous activity if they represent information about the same object. Synchronous activity occurs in the gamma band (25-60Hz) (ma

27、ximum period of about 40 msec.) frequency drifts by 5-10Hz, but synchronization stays stable for 100-300 msec In humans EEG and MEG signals exhibit power spectrum shifts consistent with synchronization of cell ensemblesorienting or investigatory behavior; delayed-match-to- sample task; visuo-spatial

28、 working memory task,Predictions: constraints on reflexive inference,gamma band activity (25-60Hz) underlies dynamic bindings (the maximum period 40 msec.) allowable jitter in synchronous firing 3 msec. lead/lag.only a small number of distinct conceptual entities can participate in an episode of rea

29、soning7 +/- 2 (40 divided by 6)as the number of entities increases beyond five, their activity starts overlapping, leading to cross-talkNote: Not a limit on the number of co-active bindings!,Predictions: Constraints on reflexive reasoning,A large number of relational instances (facts) can be co-acti

30、ve, and numerous rules can fire in parallel, but only a small number of distinct entities can serve as role-fillers in this activity only a small number of instances of the same predicate can be co-active at the same time the depth of inference is bounded systematic reasoning via binding propagation

31、 degrades to a mere spreading of activation beyond a certain depth. 2 and 3 specify limits on Shrutis working memory,Massively Parallel Inference,if gamma band activity underlies propagation of bindings each binding propagation step takes ca. 25 msec. inferring “John may be hurt” and “John may have

32、slipped” from “John fell” would take only ca. 200 msec. time required to perform inference is independent of the size of the causal model,Probabilistic interpretation of link weights,P(C/E) = P(E/C) P(C)/P(E),Encoding X-schema,Questions,Representing belief and utility in Shruti,associate utilities w

33、ith states of affairs (relational instances) encode utility facts: context sensitive memories of utilities associated with certain events or event-types propagate utility along causal structures encode actions and their consequences,Encoding “Fall = Hurt”,Fall,Hurt,Focal-clusters augmented to encode

34、 belief and utility,Attack,Behavior of augmented Shruti,Shruti reflexively Makes observations Seeks explanations Makes predictions Instantiates goals Seeks plans that enhance expected future utility identify actions that are likely to lead to desirable situations and prevent undesirable ones,Shruti

35、suggests how different sorts of knowledge may be encoded within neurally plausible networks,Entities, types and their relationships (John is a Man) Relational schemas/frames corresponding to action and event types (Falling, giving, ) Causal relations between relational schemas (If you fall you can g

36、et hurt) Taxon/Semantic facts (Children often fall) Episodic facts (John fell in the hallway on Monday) Utility facts (It is bad to be hurt), Lokendra Shastri ICSI, Berkeley,Current status of learning in Shruti,Episodic facts: A biologically grounded model of “one-shot” episodic memory formation Sha

37、stri, 1997; Proceedings of CogSci 1997 _2001; Neurocomputing _2002; Trends in Cognitive Science _In Revision; Behavioral and Brain Science (available as a Technical Report), Lokendra Shastri ICSI, Berkeley,current status of learning in Shruti,Work in Progress Causal rules Categories Relational schemas Shastri and Wendelken 2003; Neurocomputing, Lokendra Shastri ICSI, Berkeley,Questions,