What is Intelligence?

• Let’s try to define intelligence

• Definition (Merriam Webster):
– Capacity for learning, reasoning, understanding and similar forms of
mental activity
– Aptitude in grasping truths, relationships, facts, meanings
• Problem: what do reasoning and understanding mean?
What does “grasping” mean?
• Reasoning:
– Process of forming conclusions, judgments, inferences from
facts
• Understanding:
– Ability to get the meaning of and judge, to know and
comprehend
• Comprehension:
 What is Intelligence? Cont.
• Can we find a single, non-circular definition?
– Unfortunately we only have one instance of intelligence to
study: man
• How about enumerating a list of features that we think
are involved in intelligence? :
– reasoning, inferencing, problem solving
• what forms of reasoning? deduction, induction, abduction
– learning, generalization, recall, analogy
– common sense, intuition, emotion, self-awareness
• Which of these are necessary for intelligence? Which
are sufficient?
– can we have intelligence without learning?
– can we have intelligence without self-awareness and emotion?
 AI Defined
• Textbook definition:
– AI may be defined as the branch of computer science that
Thinking
is concerned with the automation of intelligent behavior
• Other definitions: machines or
– The exciting new effort to make computers think … machine
machines with minds intelligence
– The automation of activities that we associate with human
thinking (e.g., decision-making, learning…)
– The art of creating machines that perform functions that Studying
require intelligence when performed by people cognitive
– The study of mental faculties through the use of faculties
computational models
– A field of study that seeks to explain and emulate
intelligent behavior in terms of computational processes Problem
– The study of how to make programs/computers do things Solving and
that people do better CS
 Intelligence vs Intelligent Behavior
• Some of the previous definitions draw a
distinction between
– intelligence (machines that think)
– and intelligent behavior (machines that are
programmed to exhibit a behavior that looks like
intelligence
• Is there a difference?
– if so, is the difference significant?
– is intelligent behavior “good enough”?
• true machine intelligence: strong AI
• machine’s exhibiting intelligent behavior: weak AI
 Physical Symbol System Hypothesis
• A Physical Symbol System (PSS) consists of
– symbols (patterns)
– expressions (legal combinations of symbols)
– processes (to manipulate symbols and expressions into new
expressions)
• The PSS Hypothesis states that a PSS has the necessary
and sufficient means for intelligent action
– the hypothesis was first defined by Newell and Simon
– it has been proved that a PSS is Turing computable so that a
computer is a PSS
• If the PSS Hypothesis is true, then it should be possible
to program a computer to produce intelligent actions
(weak AI)
– if the PSS Hypothesis is not true, what are the consequences
for AI?
 Turing Test
• Mathematician Alan • If the interrogator
Turing devised a test cannot tell which is the
for defining artificial human and which is
intelligence: the computer, then the
– an interrogator poses computer passes the
questions to two Turing Test and should
entities, a human and a be considered
computer intelligent
– Turing first called this
the Imitation game but
has since been renamed
the Turing Test – a test
 More on the TT
• How useful is this test?
– consider the Eliza program (see the next slide)
• does it pass the Turing Test? if so, is it intelligent?
• Many programs now pass the Turing Test
– Nils Nilsson points out the problem with the PSS Hypothesis
and the TT (also reflected by Eliza and other programs):
• there is no symbol grounding – that is, the symbols of a PSS are
merely symbols to be manipulated based on rules
• the PSS has no understanding of what those symbols represent and
therefore no understanding of what conclusions might be drawn
• AI programs in general are shallow
– they recognize information syntactically (by patterns
of symbols), not semantically (by the meanings of
those patterns)
•
•
•
Person: Men are all alike.
Eliza: In what way?
Person: They’re always bugging
Eliza
us about something or other. • Eliza consists of a series of rules
• Eliza: Can you think of a specific
example? (if-then statements)
• Person: Well, my boyfriend made – sample syntactic manipulations:
me come here.
• • <X> me <Y>  <X> you <Y>?
Eliza: You’re boyfriend made you
come here? • I like <Y>  Why do you like
• Person: He says I’m depressed <Y>?
much of the time. • <X> are like <Y>  In what way?
• Eliza: I am sorry to hear you are
depressed. • <X> {mother | father | brother |
• Person: It is true, I am unhappy. sister}  Tell me more about your
• Eliza: Do you think coming here family
will help you not to be unhappy? • <X>  Can you think of a specific
• Person: I need some help, that example?
much seems certain.
• Eliza: What would it mean to you • Eliza had no understanding of the
if you got some help?
• Person: Perhaps I could learn to
text input or its own responses
get along with my mother. – try a non-sensical sentence, you will
• Eliza: Tell me more about your get a non-sensical response!
family.
 The Chinese Room Problem
• You are in a room with a book that contains pages of
Chinese symbols
– your job is to retrieve a question, written in Chinese on a piece
of paper passed into the room, look up the associated response
in the book, write down that response on a piece of paper and
pass that paper out of the room

Question (Chinese)

Storage Book of Chinese Symbols

You
 Chinese Room Continued
• The room is analogous to a computer:
– you = central processing unit
– book = program
– conveyor belt = Input/Output
– storage = memory/disk
• What do the symbols mean? Do you understand them?
– if you do not understand the Chinese symbols, can we say that
the computer understands the symbols it uses (ASCII, binary,
instructions, input, output?)
• What we see here is that a computer is a symbol
manipulating device – it follows rules (a program and
the machine’s microcode) but does not understand what
it is doing
– can there be intelligence without understanding?
– for instance, do you understand the symbols that you
manipulate (a red light for instance) or do you merely respond
to your input?
 The Consequence
• Since the Chinese Room Problem points out that a
computer probably does not understand the symbols,
should this concern us?
• Can we program a computer to be intelligent?
– how important is semantics
– that is, can we somehow ground the symbols to meaningful
information in the computer?
• Strong AI vs. Weak AI: the difference between
semantic-based programs and syntactic-based programs
– or, the difference between simulating intelligence and
performing in an intelligent way
– in the former, we try to capture intelligence in the machine
– in the latter, we merely program the computer with knowledge
and processes to apply that knowledge in a way similar to how
humans might apply the knowledge
 What does AI do?
• To some, AI means different things
• But traditionally, AI is an effort to solve problems by
applying knowledge and so we must answer these
questions:
– how do we represent knowledge
– how do we apply that knowledge
• We will examine problems such as:
– diagnosis and other forms of reasoning
– planning, design and decision making
– learning
– recognition and perception
– understanding
• often, the problems that we try to solve in AI require a lot of human
knowledge – we may need access to human experts to acquire that
knowledge and codify it
 Representations
• Consider the “mutilated chess board”
– how can you place dominoes on the mutilated chess
board so that all squares are covered?
– should we represent the chessboard visually as shown to
the right? use a 2-D array? or merely represent it like
this: 32 black squares, 30 white squares?
• Consider the game of tic-tac-toe
– data structure: 1-D array of 9 elements or 3x3 array?
– knowledge:
• we could store for each board configuration, the best move to take, this would
require 3^9 different board configurations! (table look-up approach)
• we could store rules that say, for each turn (1-9) what type of move should be
made (rule-based approach)
• we could derive a function which evaluates a board configuration for its
“goodness” and select a move based on which one is judged best (heuristic
approach)
– which approach is the most efficient?
 Table-Lookup vs. Reasoning
• In our tic-tac-toe example, we see one solution is to
have a table of all best moves
– this is impractical for most problems, consider chess or a
program like Eliza
• Instead, we want to opt for a solution that relies on
knowledge and reasoning over that knowledge
– in chess, we define rules that encapsulate chess strategies
– in diagnosis, we implement reasoning by means of “chaining”
rules that map symptoms to diseases
– in planning, we represent goals by enumerating the tasks
needed to accomplish those goals and implement reasoning by
“chaining” through the rules from goals to tasks to subtasks
 Representational Techniques
• Predicate calculus
– known items are predicates
– implication rules are used for reasoning
• Production systems
– knowledge is represented as if-then rules
– use forward or backward chaining to reason
• Graph theory
– knowledge is stored as nodes and links in a graph (or tree)
– search the graph/tree for a solution
• Semantic structures
– store knowledge as categories, instances, and their attributes
– semantic networks are a visual form, frames are the precursor of OOPLs
• Statistical/mathematical approaches
– primarily added to one of the above techniques to portray uncertainty
• Subsymbolic approaches (neural networks)
 Areas of Study
• Computer Science – algorithms, data representations,
programs to test theories
• Psychology – theories of mind, memory, learning,
experiments with human and animal intelligence
• Philosophy – mind/body problem, study of logic
• Linguistics – study of language (syntax, semantics)
• Neurology/Biology – study of the brain (both human
and animal), study of memory, learning
• Engineering – many AI domains are in engineering
disciplines, also AI is often thought of as much as
engineering as it is a science
• Mathematics – many algorithms are mathematical in
nature (neural networks, statistical approaches)
 Problem Areas
• Diagnosis
• Understanding/Recognition
– often tied in with perception
• Natural Language Processing
• Planning/design & decision making
• Game playing
• Automated theorem proving
• Learning (symbolic, subsymbolic, evolutionary)
• Agents and communication
• Ontologies and web applications
• Robotics (which combines several of the above)