A Cognitive Substrate for Human-Level Intelligence Nick Cassimatis In collaboration with Paul Bello, Magda Bugajska, Arthi Murugesan Human-Level Intelligence.

A Cognitive Substrate for Human-Level Intelligence

Nick CassimatisIn collaboration with

Paul Bello, Magda Bugajska, Arthi Murugesan

Human-Level Intelligence Laboratory

N.L. Cassimatis (2006). A Cognitive Substrate for Human-Level Intelligence. AI Magazine. Volume 27 Number 2.

General and Human-Level Intelligence

• General Intelligence• Human-level Intelligence is a useful proxy for

this.– It is a useful lower bound on what you can aim for.– If you humans can do X (e.g., use natural language),

then I am confident a computer can do X. If humans cannot do X (e.g., find prime factors of million-bit numbers), I am less confident.

– Humans are the most general intelligence we know of, so you can learn a lot about observing us (including introspection.)

• Thinking exactly like a human is not a hard constraint

Obstacles: Profusion and Integration

• Profusion of knowledge.• Profusion of algorithms.• Difficulty of integrating it all.

Great amount and variety of knowledge

Great amount and variety of knowledge

• Even very simple situations can require a much knowledge.– Knowledge about a piggy banks (Charniak):

• It is used to store money.• If you shake it and there is no sound, it is empty.• You can remove money by breaking it.• You can remove money by turning it upside down

and shaking it.• The more money you put in, the more you get out.• (dozens more pieces of knowledge).

– Exceptions to each point.• Cyc: Millions of assertions, nowhere near

complete.• How do we get all this into one computer

program?

Diversity of algorithms

• E.g., Natural language conversation– Vision for recognizing faces, tracking eyes,

gestures.• PCA, Bayesian networks, Kalman filters.

– Acoustic speech recognition.• Hidden Markov Models, Fourier Transforms.

– Syntax, phonology, morphology• Search- or table-based parsers, rules, statistical N-

gram models.– Semantics and Pragmatics (including semantics

and pragmatics)• Almost everything.

• There are often dozens or hundreds of variations within each algorithmic class.

Integration

• How do you get all these algorithms and data structures to work with each other– Procedural integration: Bayes nets,

logic theorem provers, case-based reasoning, neural networks … ?

– Knowledge integration: Scripts, frames, logical propositions, patterns of activation … ?

Learning

• Commits you to weak representations and execution algorithms because these are easier to prove theorems about and be general.

• You need rich conceptual foundation to do learning in the first place.

How do we deal with this?

• Profusion: Cognitive Substrate• Integration: Polyscheme

Cognitive substrate

• Small set of reasoning mechanisms can underlie the whole range of human cognition.

• Preliminary guess at what would be a good substrate:– Time, space, causality, identity, events,

parthood, desire.• Hypothesis: Once you have

implemented a substrate, the rest of AI is relatively simple.– Substrate is AI-complete.

Evidence for substrate

• Personal experience.• Linguistics.• Psychology.• Neuroscience.• AI.• Evolution and learning…

Evolution• We evolved to deal with a relatively immediate and

concrete physical and social world, not to – Contemplate life on Mars. – Trade stock options.– Explore number theory.– Design airplanes.– Repair speed boats.– Calculate tips.– Market insurance policies.– Etc.

• Whatever mechanisms we use to reason about these were originally designed to deal with the physical and social world.

• Hence, human social and physical reasoning mechanisms are sufficient for the full array of human reasoning.

Learning

• What is it that kids have that give them the ability to learn so much, to be so general?– Substrate mechanisms.– Mechanism for mapping.– Mechanisms for learning.– Mechanisms for being taught.

Substrate research

• Overall approach:– Build substrate (2-4 year old?)– Turn it loose on the world.

• Building the substrate– First guess (physical reasoning)– Map onto several domains (epistemic

reasoning, syntax, word learning)– Each mapping leads to refinements and

generalizations– Learning mechanisms (analogy)

Contrast

• Many people dream of building a baby and setting it loose on the world.

• Contrast– Need a richer substrate.– Need to integrate learning with

reasoning.

Building a substrate

• Reasoning about time, space, causality, identity, events, parthood, desire …– Requires integration of temporal,

spatial, causal … data structures and algorithms.

• Polyscheme is an approach to this problem.

Common Functions

• Basic functions: – Forward inference.– Subgoaling.– Identity matching.– Representing alternate worlds.

• Basic functions can be computed using different representations:– E.g., subgoaling:

• Logic: when B H and want to know if H, make a subgoal of B.

• Neural Network: To know the value of the output units, make a goal of the input units.

• Perception: To know what is at P, point the camera to P.

AI algorithms are ways of ordering common functions

• Counterfactual reasoning– When uncertain about A, simulate the world where A and

simulate the world where not-A.• Backtracking search

– Nested counterfactual reasoning.• Stochastic simulation

– When you think A is more likely than not-A, simulate the world where A is true more often than the world where A is not true.

• Logic-theorem proving– When uncertain about P

• Ground P if you can.• Subgoal on P if you can.

• Means-ends planning.– When you want G, and A achieves G,

• Simulate the world where A is true and subgoal on A.

Integration of algorithms

Integration of representations

Physical reasoner demonstrates flexible

integrationReactive/Deliberative Robot architecture– Combines means-ends

planning, logical inference, production rules, neural networks, truth maintenance, reactive subsystem etc.

Promising approach to (hard) substrate problems.– In physical reasoner.– Adding algorithms and

representations adds to huge increase in efficiency.

Several problems mapped onto physical reasoning substrate.

N. L. Cassimatis, J. Trafton, M. Bugajska, A. Schultz (2004). Integrating Cognition, Perception and Action through Mental Simulation in Robots. Journal of Robotics and Autonomous Systems. Volume 49, Issues 1-2, 30 November 2004, Pages 13-23.

Example: SyntaxMurugesan, N.L. Cassimatis (2006). A Model of Syntactic Parsing Based on Domain-General Cognitive Mechanisms. In Proceedings of 28th

Annual Conference of the Cognitive Science Society.

N. L. Cassimatis (2004). Grammatical Processing Using the Mechanisms of Physical Inferences. In Proceedings of the Twentieth-Sixth Annual Conference of the Cognitive Science Society.

Show how to map syntactic parsing to physical reasoning.• What could words, phrases, case, empty categories, traces,

long-distance dependencies, coreference, subjacency, anaphora, etc. have to do with gravity and collision?

• If these two domains have underlying unity, then you cannot quickly rule out mappings between other domains.

SyntaxVerbal World Physical World

World, phrase, sentence Event

Constituency Parthood

Phrase structure constraints Physical constraints

Word/phrase category Categories

Word/phrase order Temporal order

Phrase attachment Event identity

Coreference/binding Object identity

Traces Object permanence

Short- and long-distance dependencies

Apparent motion and long paths.

Syntax

Word Learning

• One-shot, non-associative word learning

M. Bugajska, N.L. Cassimatis (2006). Beyond Association: Social Cognition in Word Learning. In Proceedings of the International Conference on Development and Learning.

Theory of Mind

• Use counterfactual and default reasoning mechanisms to reason about other people’s beliefs.

P. Bello & N.L. Cassimatis (2006). Developmental Accounts of Theory-of-Mind Acquisition: Achieving Clarity via Computational Cognitive Modeling. In Proceedings of 28th Annual Conference of the Cognitive Science Society.

P. Bello & N.L. Cassimatis (2006). Understanding other Minds: A Cognitive Modeling Approach. In Proceedings of the 7th International Conference on Cognitive Modeling.

Summary of progress

• Preliminary implementation (physical reasoning)– Demonstrates Polyscheme enables advance in flexibility,

integration and power of intelligent systems.

• Manually mapped onto several domains (epistemic reasoning, syntax, word learning, wargaming)– Each mapping demonstrates the plausibility of the

substrate appraoch.• Each mapping leads to refinements, generalizations and

eliminations about the substrate.

• Learning mechanisms (analogy).– Just starting

• Teaching the substrate (this will gradually result from our NLP work).

What this demonstrates

• Cognitive substrate enables a real advance towards solving the profusion and integration problem.

• It enables qualitative advances in capabilities of intelligent systems.

• It enables faster development of systems.

Future work

• Keep doing mappings.– Pragmatics.– Metacognition.– Self-awareness, consciousness.

• Use insights from this to enhance substrate.• Automate mappings.• Keep driving this process towards the goal

of having a 2-4 year old intelligence that can learn from interacting with the world and people.

How people can help

• Software engineering. • Find a domain and do a mapping.• Add an algorithm or subdomain to

the substrate.