1. A Cognitive Substrate for Achieving Human-Level Intelligence Nicholas L. Cassimatis AI Magazine Vol. 27, No. 2 (2006) Summarized by Eun Seok Lee BI. 2008. 04. 07.

2. A Profusion Problem “When is the first Pittsburgh Steelers game after (the World Series, Thanksgiving, my daughter’s birthday, the next full moon)?” Cf. Cyc (Lenat and Guha 1990) & ThoughtTreasure (Mueller 1998) – not yet to provide a comprehensive store of all the knowledge for such queries. Question knowledge info National holidays User info Celestial movements

3. The Procedural Profusion Problem A human utterance hidden Markov models Convert continuous acoustic signal into discrete representation of the phonemes, morphemes, and words chart- or search-based algorithms Identify the syntactic structure logical, case-based, and probabilistic methods Reasoning about the world and other people’s intention

4. The Procedural Profusion Problem Too many computational problems involved with difficult-to- integrate methods. Each aspects can best be dealt with different methods Tanenhaus and Trueswell 1995 – Above is very difficult problem – only multiply for systems that integrate more of human-level intelligence. Then what about enabling computers automatically storing and collecting knowledge and algorithms? – not sufficient; only learn to delegate among existing algorithms. And, control and data structures of these different classes of algorithms are very difficult to integrate Thus, the profusion problem is a genuinely difficult integration problem.

5. The Cognitive Substrate Hypothesis Human’s earlier cognitive mechanism with adapting mechanism is sufficient to achieve human-level intelligence in all domains. A relative small set of computational problems that once solved, (“cognitive substrate,”) can be adapted to solve other problems.

6. An Example of a ‘Cognitive Substrate’ Basic social, physical reasoning problem Reasoning Temporal intervals Causal relations Desire Objects Events Belief Ontologies “Cognitive substrate” – A set of computational mechanisms i.e. “Once a set of ‘cognitive substrate’ is constructed, the rest will be relatively easy.”

7. The Benefits of the Cognitive Substrate Hypothesis 1.Smaller problem 2.Quicker intelligent system development 3.Easier integration across domains If cognitive substrate underlie cognition in most domains of human cognition, then much of the problem of human-level cognition becomes more tractable. AI, Cognitive psychology, linguistics, neuroscience support the cognitive substrate hypothesis.

8. Implicit Substrate Hypothesis in Much AI Research Representative research: –Search through a state space (Newell and Simon 1972) –Updating probabilities in a Bayesian network (Pearl 1988) –A modest set of primitives that can represent much of the semantics of human language (Shank 1975) – not implemented because reasoning problems with these primitives are not yet achieved. –Thus require benefits of each specific class of AI methods should be integrated into one system.

9. Linguistic Semantics Jackendoff – structures used to represent the semantics of a relatively small set of semantic fields can be used to represent the semantics of many other semantic fields. i.e. primitives such as cause, go, path, to, from are common in other frameworks. The formalization: –“John entered the room.” -> GO (John, [path [to: room]]) –“John left the room.” -> GO (John, [path [from: room]]) Large set of word classes with only a few more primitives – support the notion that a relatively small set of mechanisms can lead to human-level intelligence in all domains.

10. Cognitive Psychology and Neuroscience Much nonspatial or physical thought involves mechanisms which are for spatial and physical ones. Barsalou et al (2003) – human mapping visual and motor representations onto abstract, nonphysical concepts Spivey et al (2001) – associating sitting position with verb ‘push/respect’ Ricahrdson et al (2003) – harder understanding sentences of certain combination of images and words. Warington et al (1984) – visual and motor regions activation during nonperceptual cognition

11. Constituents of Cognitive Substrates Cassimatis 2002 – preliminary lists includes reasoning about time, space, part-hood, categories, causation, uncertainty, belief, desire. Implied research program –1) identify and implement a cognitive substrate –2) find mappings from multiple domains onto the cognitive substrate –3) automate the process of adapting a cognitive substrate so that it can solve problems in other domains

12. Implementing a Cognitive Substrate Polyscheme cognitive architecture: –enables multiple computational methods to be implemented much more ubiquitous. Two principles: –CFP common function principle –MIP multiple implementation principle Attention Fixation –Very different algorithms can all be reformulated in terms of sequences of attention fixations

13. Common Functions Forward inference. Subgoaling. Simulate alternate worlds. Identity matching. i.e.implemeting an algorithms –Search. “When uncertain about whether A is true, represent the world where A is true, perform forward inference, represent the world where A is not true, perform forward inference. If forward inference leads to further uncertainty, repeat.” –Stochastic simulation. “When A is more likely than not-A, represent the world where A is true and perform forward inference in it more often than you do for the world where not-A is true.”

14. MIP for forward inference Neural Networks. –The activation of input units of a feedforward neural network leads to a change in the activation of the output units of the network. –Rrepresent facts that can be inferred from the facts represented by the input units. Forward rule changing. –Production systems can be constructed to match the left-hand sides of production rules against a set of currently known facts to infer new facts represented by the right hand sides of rules. Ontologies. –When an object, o, is a member of category C in a category hierarchy, one can infer that o is a member of C1 … Cn, the ancestors of C.

15. Implementing an algorithm A CFP Subgoaling Simulate alternate worlds Search Stochastic simulation Identity matching Forward inference Algorithms Neural Networks Forward Rule Chaining Ontologies An MFP

16. Attention Fixation Very different algorithms –can all be reformulated in terms of sequences of attention fixations Inference algorithms (originally based on very different computational formalisms): –can be executed as sequences of a small set of common functions (according to the CFP) that can be easily interleaved –and that these common functions can be implemented using many different algorithms (according to the MIP)

18. Leveraging a Cognitive Substrate How a substrate on a Polyscheme model of human physical reasoning used to construct a natural language parser? Many grammatical structures have analogues to nonlinguistic cognitive structures. Formalism: –Events (e), Objects (o), Places (p) –Category(e, MotionEvent), Agent(e, x), Origin(e, p1), Destination(e, p2), Occurs(e, t) –An unsupported object falls is: Location(o, p1, t1) + Below(p2, p1) + Empty(p2, t1) Category(e, MotionEvent) + Origin(e, p1) + Destination(e, p2) + Occurs(e, t2) + Meets(t1, t2).

19. Leveraging a Cognitive Substrate Utterances are events: Category(e, dog-utterance), Occurs(e, t). Word order is temporal order: Category(e1, JohnUtterance), Occurs(e1, t1) Category(e2, BitUtterance), Occurs(e2, t2) Meets(t1, t2) … Physical and linguistic events both belong to categories organized hierarchically –Constituency is a parthood relation. –Coreference and binding are object-identity relationships. –Phrase attachment is an event identity relationship.

21. Conclusion: Benefits of Cognitive Substrates 1.Much easier to create an intelligent system for new domains 2.Much easier integration among domains 3.The problem of achieving human- level AI is reduced and simplified by mapping a relatively small set of problems onto a substrate