Intelligence: Real and Artificial

Download Report

Transcript Intelligence: Real and Artificial

Computers & Thought
Lecture 1
January 5th, 1999
CS250
Lecture 1
CS250: Intro to AI/Lisp
What is cognition?
• Cognition is widely studied: philosophy,
psychology, and other fields
• Can we implement computer programs
that think?
– Model the process
– Create the result
– Example: airplane flight
Lecture 1
CS250: Intro to AI/Lisp
Is cognition computation?
• Computation is what you can do with a
Turing machine (Church-Turing)
– What's a Turing machine?
– Model for a Turing machine?
• Need states and operations
Alan Turing
Copyright (c) 1997. Maxfield &
Montrose Interactive Inc.
– Brain states
– Operations that move among states
Lecture 1
CS250: Intro to AI/Lisp
The Turing Test
• How to tell if a computer is intelligent?
• Use the Turing test
– Boston Computer Museum 1991 test
• Is this a good definition of intelligence?
Lecture 1
CS250: Intro to AI/Lisp
Is Computation Enough for
Cognition?
• Chinese Room Argument
– Does the man in the room understand
Chinese?
– Does it matter?
• What's the difference between a native
Chinese speaker, and the "room in a
man"?
Lecture 1
CS250: Intro to AI/Lisp
Refutations of the Chinese Room
Argument
• Systems reply
• Another reply:
– Searle argues
1) Some elements don't understand Chinese
(human, paper and the book)
2) Elements that don't understand cannot be
pieced together to create a whole that does
understand
Lecture 1
CS250: Intro to AI/Lisp
What Do You Need to Pass a
Turing Test?
• Conversational skills (known as natural
language processing, or simply NLP)
• Store of knowledge
• Reasoning
• Learning
Lecture 1
CS250: Intro to AI/Lisp
Approaching AI
• Building a brain in a computer
– Cognitive modeling
• High-school geometry approach
– Logic & inference
• Agent approach
– Rationality
• Experiential approach
– Case-based reasoning
Lecture 1
CS250: Intro to AI/Lisp
Other Fields & AI
• Philosophy
– Questions of brain and mind
– Consciousness
– Logic
Lecture 1
CS250: Intro to AI/Lisp
Mathematics
• Logic
• Theory of computability
• Probability
Lecture 1
CS250: Intro to AI/Lisp
Logic in AI
• Inference
– Possible reasoning process for cognition
• Representation
– Logical statements
– Observations about the world
– States of the world
Lecture 1
CS250: Intro to AI/Lisp
How hard is hard?
• Computability / complexity theory
• How hard is a problem?
Lecture 1
CS250: Intro to AI/Lisp
"What are the odds?"
• World is full of uncertainty
• Probability helps formalize uncertainty
• Bayes Theorem is key:
From http://members.tripod.com/~Probability/bayes01.htm
Lecture 1
CS250: Intro to AI/Lisp
Psychology
• "Information processing view"
– Response to behaviorism
Lecture 1
CS250: Intro to AI/Lisp
Computer Engineering
• Sets AI apart from non-computer
disciplines
• Focus on implementation
Lecture 1
CS250: Intro to AI/Lisp
Linguistics
• Link between language and thought
– MITECS article on Language of Thought
Lecture 1
CS250: Intro to AI/Lisp
Sapir-Whorf hypothesis
• Linguistic Relativity
– Structural differences between languages
will generally be paralleled by nonlinguistic
cognitive differences
• Linguistic determinism
– Structure of a language strongly influences
or fully determines the way its native
speakers perceive and reason about the
world
Lecture 1
CS250: Intro to AI/Lisp
What about Sapir-Whorf?
• Anthropologist John Lucy
– Speakers of languages with different basic
color vocabularies might sort non-primary
colors (e.g., turquoise, chartruese) in
slightly different ways
• Psychologist Alfred Bloom's claim
– No distinct counterfactual marker in
Chinese --> difficult for Chinese speakers
to think counterfactually
Lecture 1
CS250: Intro to AI/Lisp
Cognition and Language
• Abilities to learn and use language part
of our general intelligence
– Language Specific Impairments
– Williams syndrome
Cognition and language can be decoupled
Lecture 1
CS250: Intro to AI/Lisp
Natural Language Processing
• NLP is one of the most difficult tasks in
AI (AI-complete)
• Why?
– Ambiguity resolved by context
– Computers lack context
Lecture 1
CS250: Intro to AI/Lisp
Early AI
• 1952-1956: Samuel's checkers playing
program beats Samuel
• Summer 1956 @ Dartmouth: Summer
workshop with John McCarthy Marvin
Minsky, Claude Shannon and others
• 1958
– Lisp
– Time sharing
– Advice Taker
Lecture 1
CS250: Intro to AI/Lisp
Knowledge-Based Systems
• Late 60's, early 70's: Big talk falls flat
• Knowledge-based systems
– Know about the world
• DENDRAL
– Deduced molecular structure from mass
spectrometry
– Encoded rules from experts
• MYCIN
Lecture 1
CS250: Intro to AI/Lisp
Linguistics in AI
• The "Yale school"
– Roger Schank jumped ship from linguistics
to AI
– "There's no such thing as syntax"
• What do you need to understand
language?
– Heavy on domain knowledge
– Scripts, CBR
Lecture 1
CS250: Intro to AI/Lisp
Representational Systems
• Driven by big problems
– Battlefield communication
– Logistics
– Campaign planning
• Scaling up
– Prolog for rules
– Frames (from MM) for structured
representations
Lecture 1
CS250: Intro to AI/Lisp
Recent History
• Assist instead of replace
• Neural networks are back
– Perceptrons, by Minsky and Papert
– Backpropagation brings NN's back
• Probabilistic systems
– Bayesian networks (e.g., Koller & Horvitz)
• Information retrieval and analysis
Lecture 1
CS250: Intro to AI/Lisp
Intelligent Agents
Tasks or Problems
Action selection and planning
Agent communication languages
Agents in entertainment applications
Believable agents
Collaboration between people and agents
Communication between people and agents
Coordinating perception, thought, or action
Expert assistants
Information agents
Integrated theories of intelligence
Knowledge acquisition and accumulation
Learning and adaptation
Modeling emotion
Multi-agent communication, coordination,
or collaboration
Multi-agent simulation
Multi-agent teams
Techniques or Algorithms
Algorithms for negotiation
Artificial market systems
Cognitive models
Evaluations and implemented systems
Game-theoretic modeling of the behavior of other
agents
Logic-based agent communication languages
Meta-modeling of an agent by itself
Mobile agents
Task-specific agent architectures
Intelligent Interfaces
Tasks or Problems
Auditory scene analysis
Computer-aided instruction
Conversation
Help desks
Intelligent buildings or rooms
Models of human speech perception
Music perception
Speech coding
Speech recognition
Speech synthesis
Techniques or Algorithms
Evaluations and implemented systems
Hidden Markov models
Learning interaction models
Learning user preferences
Student modeling techniques
Knowledge Representation and Reasoning
Tasks or Problems
Causal reasoning
Common-sense reasoning
Constraint satisfaction tasks
Design, modeling, simulation, or diagnosis
Game playing
Reasoning about embedded systems
Reasoning about relevance
Representations of belief, intention, time, space,
action, or events
Spatial and geometric reasoning
Temporal reasoning
Lecture 1
Techniques or
Algorithms
Analogical reasoning
Boolean satisfiability
Case-based reasoning
Complexity analysis
Constraint satisfaction
Description logics
Design, analysis or evaluation of ontologies
Design and evaluation of implemented KR systems
Game-playing methods
Genetic algorithms
Integer and constraint programming
Logic programming and theorem proving
Modal logics
Model-based reasoning
Parallel and distributed implementations
Qualitative reasoning
Search or optimization
Significant applications
Simulated annealing
Temporal logics
Machine Learning and Discovery
Tasks or Problems
Abstraction learning
Active learning
Computational learning theory
Constructive induction
Data mining
Learning and planning
Learning dynamics
Learning in computational biology
Learning in embedded systems
Learning in information retrieval
Learning on the Internet
Online learning
Reinforcement learning
Scientific discovery
Speedup learning
Supervised learning
Theory refinement
Unsupervised learning
Techniques or Algorithms
Case-based learning
Comparative analyses
Decision-tree learning
Empirical evaluation of learning algorithms
Evaluations and implemented systems
Evolutionary computation
Genetic programming
Inductive logic programming
Learning algorithms with provable properties
Learning belief networks
Learning mixture models
Multi-strategy learning
Neural nets
PAC learning and beyond
Reinforcement learning algorithms
Specialized learning algorithms
Theory of model selection and evaluation
Natural Language Processing
Tasks or Problems
Dialog
Discourse
Generation
Information extraction from the Web
Machine translation
Multimedia models
Understanding
Techniques or Algorithms
Evaluations and implemented systems
Hidden Markov models
Statistical or corpus based methods
Planning, Scheduling and Control
Tasks or Problems
Active perception and sensor-based
planning
Agent architectures for planning and
control
Decision-theoretic planning
Mixed-initiative planning
Multi-agent planning
Plan and schedule visualization
Plan execution, monitoring or replanning
Planning and learning
Resource management
Scheduling
Techniques or Algorithms
Comparative analyses
Compilation to SAT
Constraint management approaches
Discrete control theory approaches
Empirical evaluations
Evaluations and Implemented systems
Fuzzy control techniques
Graphplan-based algorithms
MDP planning
Partial-order planning
Planning using dynamic belief networks
Scheduling algorithms
Specialized planning algorithms
Robotics
Tasks or Problems
Behavioral control
Dynamical control systems
Geometric motion planning
Human robot interaction
Mapping and exploration
Micro-robotics
Mobile robotics
Multi-robot coordination
Robot control architectures
Robot learning
Techniques or Algorithms
Coordination methods for multirobot
systems
Evaluations and implemented systems
Fuzzy logic controllers
Pomdp localization methods
Subsumption architecture
CS250: Intro to AI/Lisp
Uncertainty in AI
Tasks or Problems
Computation and action under bounded resources
Control of computational processes
Decision making under uncertainty
Decision-theoretic planning and reasoning
Diagnosis: medical, mechanical, or software
Enhancing the human-computer interface
Integration of logical and probabilistic inference
Learning and data mining
Stochastic modeling
Temporal reasoning
Uncertain reasoning in embedded systems
Techniques or Algorithms
Abstraction in representation and inference
Algorithms for learning and data mining
Automated construction of decision models
Automated explanation of results
Beyond Markov models
Comparative analyses of algorithms and systems
Design and performance of architectures for real-time
reasoning
Discovery of causal relationships
Economic models of problem solving
Empirical validation of methods
Evaluations and implemented systems
Experience with knowledge-acquisition methods
Formal languages
Game-theoretic modeling
Hybrid techniques
MDPs and Bayesian Networks
Qualitative methods and models
Representing causality
Specialized reasoning techniques
Specialized representations
Statistical methods
Time-dependent utility functions
Vision
Tasks or Problems
Active perception
Analysis of medical images
Face recognition
Hand-eye coordination
Image and video compression
Image processing
Object recognition
Perception and learning
Psychophysical modeling
Visual recognition and tracking
Visual scene analysis
Techniques or Algorithms
Evaluations and implemented systems
Markov Random fields
Neural net algorithms
Optical flow techniques
Conclusions
• AI is a young field with a tumultuous
past
• Interdisciplinary
• Humans are really, really smart
Lecture 1
CS250: Intro to AI/Lisp