Traditional questions that have been
analysed by philosophers, psychologists, and linguists:
What is knowledge?
What do people have inside their
head when they know something?
Is knowledge expressed in words?
If so, how could one know things
that are easier to do than to say, like tying a shoestring or hitting a
baseball?
If knowledge is not expressed in
words, how can it be transmitted in language?
How is knowledge related to the world?
What are the relationships between
the external world, knowledge in the head, and the language used to express
knowledge about the world?
With the advent of computers, the
questions addressed by the field of artificial intelligence (AI):
Can knowledge be programmed in a
digital computer?
Can computers encode and decode that
knowledge in ordinary language?
Can they use it to interact with
people and with other computer systems
in a more flexible or helpful way?
Artificial Intelligence raises the
same issue about knowledge and its relationship to language and to the
world that have been addressed by philosophers for the past two and a half
millennia.
1.2.
Knowledge and Models
Knowledge is more than a static encoding
of facts, it also includes the ability
to use those facts in interacting with the world.
Basic premise of AI is that knowledge
of something is the ability to form
a mental model that accurately represents the thing as well as the
actions than can be performed by
it and on it.
By testing actions on the model,
a person (or robot) can predict what is likely
to happen in the real world.
To test possible actions, AI systems
construct microworlds.
The hypothesis that people understand
the world by building mental models
raises fundamental issues for all the fields of cognitive science:
Psychology - How are models
represented in the brain, how do they
interact with the mechanisms for perception, memory, and learning,
and how do they affect or control behaviour?
Linguistics - What is the
relationship between a word, the object it
names, and a mental model ? What are the rules of syntax and semantics
that relate models to sentences?
Philosophy - What is the relationship
between knowledge, meaning,
and mental models ? How are the models used in reasoning,
and how is such reasoning related to formal logic?
Computer Science - How can
a person's model of the word be reflected
in a computer system ? What languages and tools are needed
to describe such models and relate them to outside systems?
Can the models support a computer interface that people
would find easy to use?
1.3.
Psychological Issues
Associationism:
the oldest theory of psychology;
started with Aristotle;
a sensation is associated
with an idea, and that idea leads to
another idea, which leads to still other ideas.
Behaviourists:
eliminated all talk about ideas,
mental states and thinking;
they maintain that a theory should
relate external stimulus to
observable response without any assumptions about mental
states and processes;
as an experimental technique, they
developed conditioning and
reinforcement for building and strengthening stimulus- response
chains@
.
Some comments:
conditioning cannot explain the students'
novel behaviour in analysing
the situation, predicting Minsky's responses, and planning
a strategy for conditioning him.
Language is also beyond the scope
of behaviourism - one sentence
reversed the effect of an hour of conditioning.
can only explain habitual behaviour,
cannot explain how language
can exert a powerful effect with a single sentence or even
just one word.
Conclusion
Behaviourism narrowed the scope of
psychology to such an extend
that the most interesting questions could not be asked.
Cognitive psychologists talk about
mind, intelligence, thought and knowledge.
Experiment 1: Norwegian white rat
run the maze
behaviourist would say that when
the rat learns to run a maze, the
passageways are stimuli that trigger running motions in the learned
directions. Note that
when the maze is flooded, the rat will swim the maze correctly
even though it has never associated swimming motions with
the stimuli.
cognitive psychologist (eg. Tolman
(1932)) maintains that a rat does
not respond blindly to the immediate stimulus, instead, it has
a cognitive map that relates the local surroundings to the eventual
goal.
Experiment 2: Connie happens to be
hungry when she sees a street vendor
selling ice cream. She may then walk up to the vendor,
take out money, buy some ice cream, and eat it.
Somehow, the possibility of eating
ice cream in the future "causes"
her to carry out actions in the present. But, basic laws of physics
say that future events cannot affect the present.
Behaviorists would say that the stimulus
of seeing the vendor, enhanced
by Connie's hunger, triggers a conditioned response that
leads to eating ice cream. This explanation may be right for habitual
reactions, but what about novel situations for which they
have no ready-made responses.
Cognitive psychologists would say
that when Connie sees the vendor,
she forms a model of the situation. But she also forms models
of future states where she may be eating ice cream, dining
at a restaurant, or going hungry. Which course of action she
chooses depends on her options for transforming a model of the
current state into each of the possible models. Her actions, therefore,
are not caused by future events, but by operations on models
that exist in her brain at the present.
Craik (1943) suggested, reasoning
is a system of artificial causation that
transforms models in the head.
Otto Selz (1913, 1922) developed
his theory of schematic anticipation: the solution to a problem
is not found by undirected association, but by finding
the concepts to fill in the gaps of a partially completed schema.
Indirectly, Selz described mechanisms
that were later developed for AI: backtracking,
pattern-directed invocation, and networks of concepts and
relations.
Even though behaviourism is on a
downward trend, one phenomenon, imagery#,
has remained controversial.
Psychologists (eg. Kosslyn (1980))
developed experiments that show the
importance of both image-based reasoning and conceptual reasoning:
mental images are projected on a
visual buffer. They can be scanned,
rotated, enlarged, or reduced.
novel images can be constructed from
a verbal suggestions:
Imagine George Washington
slapping Mr. Peanut on the back.
reasoning about sizes, shapes, and
actions is faster and more accurate
in terms of images.
abstract thought and logical deduction
are faster and more accurate
in terms of concepts.
a complete theory of human thinking
must show how images are interpreted
in concepts and how concepts can give rise to images.
1.4.
Linguistic Issues
Language, a means of communication
is organised in a system of complex
level of rules, each level handles one aspect of a communication
process:
Syntax studies the grammar
rules for expressing meaning in a string
of words;
Semantics is the study of
meaning itself;
Pragmatics studies how the
basic meaning is related to the current
context and the listener's expectations.
Traditional grammar consists
of informal rules that are taught in schools.
Transformational Grammar (Noam
Chomsky) is a formal theory of syntax,
but it largely neglects semantics and pragmatics. Thus, it has been
criticised as an unlikely model of how people use language.
In defence, Chomsky distinguished
competence from performance*.
He maintains that transformational grammar is an abstract theory of
competence and should not be judged
as a theory of performance.
AI needs a theory of performance
that could support communication between
people and machines.
In AI systems, conceptual graphs
are widely used for representing meaning.
Conceptual graphs emphasise
semantics.
In linguistics, Lucien Tesniere (1959)
used similar graph for his dependency grammar.
The earliest form implemented on
a computer were the correlationalnets
by Silvio Ceccato (1961).
Under various names, such semantic
nets, conceptual dependency graphs, partitioned nets, and structured
inheritance nets, the graphs have
been implemented in many AI systems.
Chomsky's students diverged from
the master's path, due to disagreement
over several issues:
roles of syntax and semantics in
generating sentences;
nature of the underlying base structure;
logic, quantifiers, and methods of
binding pronouns to their antecedents;
constraints that limit transformations
to just those patterns that actually
occur in natural languages.
Sgall (1964) proposed generative
semantics: semantic rules generate the
base structure, syntactic rules map the base into the surface structure
of a sentence, and phonological rules map the surface structure
into actual speech.
Jackendoff (1972) maintained that
different aspects of meaning are contained
in separate semantic structures. As a sentence is generated, transformations
combine the separate aspects into a single utterance.
Similar arguments were raised with
conceptual graphs.
Woods (1975) believed that the graph
should contain all the information
present in the sentence.
Like Jackendoff, Quillian maintained
that the basic meaning is separate
from the "stage direction" that determine how the meaning is expressed.
Note: The semantic base depends on
what the speaker knows about the topic.
The way the speaker presents the topic depends on pragmatics - context,
external circumstances, and the listener's expectations.
There is no reason to believe that
all these aspects of meaning originate
in a single base structure.
A sentence is derived from six different
kinds of information:
Conceptual graphs are the
logical forms that state relationships between persons, things, attributes,
and events.
Tense and modality
describe how conceptual graphs relate to the real
world. They state whether something has happened, can happen,
will happen, or should happen.
Presupposition is the background
information that the speaker and
the listener tacitly assume.
Focus is the new point that
the speaker is trying to make.
Coreference links show which
concepts refer to the same entities. In
a sentence, these links are expressed as pronouns and other anaphoric
references.
Emotional connotations are
determined by associations in the mind
of the speaker and listener.
1.5.
Intensions and Extensions
Tulving (1972) classified memories
in two categories: episodic and semantic.
Episodic memory stores detailed
facts about individual things and events
- corresponds to history and biography.
Semantic memory stores universal
principles - corresponds to dictionary
definitions.
These two categories of meaning reflect
two aspects of word meaning:
The intension of a word is
that part of meaning that follows from general
principles in semantic memory.
The extension of a word is
the set of all existing thing to which the
word applies.
The intension of mammal, for example,
is a definition, such as "warm- blooded
animal, vertebrate, having hair and secreting milk for nourishing its young";
the extension is the set of all mammals in the world.
Perception maps extensional objects
to intentional concepts and speech maps
concepts to words.
Aristotle's distinction (above) was
codified as a meaning triangle by Ogden
and Richards (1923).
The left corner is the symbol or
word; the peak is the concept, intension,
thought, idea, or sense; and the right corner is the referent, object,
or extension.
1.6.
Primitives and Prototypes
The intension of a complex concept
may be defined in terms of more primitive concepts.
Aristotle defined the concept type
MAN in terms of RATIONAL and ANIMAL.
The type ANIMAL is the genus or general type, and RATIONAL
is the differentia that distinguishes MAN from other types of
ANIMAL.
RATIONAL and ANIMAL can themselves
be defined in terms of still more primitive genera with appropriate differentiae
until, perhaps, everything
would be defined in terms of indivisible primitives.
Aristotle's primitives (also called
categories) include Substance, Quantity,
Relation, Time, Position, State, Activity, and Passivity.
Aristotle listed different categories
in different writing, but never gave a
final definitive set of primitives.
Wittgenstein(1921) stated that compound
propositions are made up of elementary
propositions, which in turn are related to atomic facts about
elementary objects in the world.
Wittgenstein (1953) repudiated his
earlier position, because concepts like
GAME has no differentiae that distinguishes games from all other
activities. Instead games share a
sort of family resemblance.
Biological classification (another
science founded by Aristotle) developed
a form of definition that does not depend on primitives.
Each species is defined by describing
a typical member, and each genus by
describing a typical species.
Mill (1965) dropped the assumption
of necessary and sufficient conditions,
but he still assumed that types were defined by primitives. He leaned towards
a probabilistic view that require a preponderance of
defining characteristics, though not necessarily all of them.
In summary, three views on definition:
Classical
A concept is defined by a genus
or supertype and a set of necessary
and sufficient conditions that differentiate it from other
species of the same genus. This approach was first stated by
Aristotle and is still used in formal treatment of mathematics and
logic. Defended vigorously by Wittgenstein earlier, then rejected
it.
Probabilistic
A concept is defined by a collection
of features and everything that
has a preponderance of those features is an instance of that concept.
This is the position taken by J.S. Mill. It is also the basis for
modern techniques for cluster analysis.
Prototype
A concept is defined by an example
or prototype. An object is an instance
of a concept c if it resembles the characteristic prototype
of c more closely than the
prototype of concepts other than c. This is
the position taken by Whewell and is closely related to Wittgenstein's
notion of family resemblances.
Zadeh (1974) tried to formalise the
probabilistic point of view, in his fuzzy
set theory.
This course adopts a compromise between
Aristotle and Wittgenstein.
1.7.
Symbolic Logic and Common Sense
From the time of Aristotle to the
19th century, logic was used to characterise
forms of reasoning in ordinary thought and language.
Boole (1854) called his rules the
laws of thought.
Frege (1879), who invented the first
complete theory of first-order logic, called his notation, concept writing.
Whitehead and Russell (1910) codified
symbolic logic in its present form as
a system of reducing mathematics to logic.
Several differences between symbolic
logic and natural language:
Interpreting v
and -> as
equivalent to English conjunction (or) and if-then.
In English, If it rains, you'll be wet is normal because there is a clausal
connection between the clauses.
In standard logic, truth of a compound
proposition depends only on the truth
of its parts, not on their meaning.
Thus, all the following statements
are true:
Either Caesar died or the moon is
made of green cheese.
If Socrates is monkey, then Socrates
is human.
If elephants have wings, then 2+2=5.
Extensionality of symbolic logic.
The English statement Every unicorn is a cow is obviously false by the
intensions of UNICORN and COW.
But in symbolic logic, that statement
is represented by the formula
"x(UNICORN(x)
--> COW(x))
which reads:
For all x, if x is a unicorn, then x is a cow.
The above formula is equivalent to:
~$x(UNICORN(x)
^ ~COW(x))
which reads:
It is false that there exist an x that is a unicorn and not a cow.
Since no unicorns exist, the statement is considered true. In English,
the intensions of UNICORN and COW make the statement false, but in symbolic
logic, the empty extension of UNICORN makes if true.
Deductive reasoning.
In logic, a proof is a sequence of formulas that starts with axioms and
generates each formula from preceding ones by manipulating symbols. When
people follow an argument, they get at its "meaning" without generating
a formal proof.
Syntax of formulas and the use of variables.
Consider the English statement and its translation into logic:
Some girl screamed. $x(GIRL(x) ^ SCREAMED(x)).
A variable is a kind of pronoun. What is unnatural is the translation of
a sentence with no pronouns into one with three.
Because the forms of symbolic logic are so different from natural language,
many people in AI rejected logic in favour of informal methods for common
sense reasoning. To explain common sense reasoning, Craik(1943) viewed
the brain as a system for making models. Refer to pp. 19.
To simulate such a system, Minsky (1975) proposed the notion of frame,
which are prefabricated patterns, assembled to form mental models. In story
understanding, if the frames do not fit together, the story is self-contradictory;
if no frames are available, the story is incomprehensible; if more than
one frame can be applied, the story is ambiguous.
To meet the objections to standard logic, conceptual graphs have been designed
as a more natural notation for logic.
1.8.
Artificial Intelligence
Artificial Intelligence is the study
of knowledge representation and their
use in language, reasoning, learning, and problem solving.
AI programs gain flexibility over
conventional systems by using a changing
knowledge base rather than a fixed, pre-programmed algorithms.
Scruffies vs Neats
In the late 70's and early 80's
the debate between the scruffies, led by Roger
Schank and Ed. Feigenbaum, and the neats, led by Nils Nilsson:
The neats argue that no education
in AI was complete without a strong
theoretical component, containing, for instance, courses in predicate logic
and automata theory.
The scruffies maintain that such
a theoretical component was unnecessary,
and harmful...
The end product of the scruffy researchers
is a working computer program,
whereas the neat researcher is not satisfied until he has
abstracted a theory from the program.
The neat view of AI assumes that
a few elegant principles underlie
all the manifestations of human intelligence. Discovery of
those principles would provide the key to the working of the
mind.
The scruffy view is that intelligence
is a kludge: people have so many
ad hoc approaches to so many different activities that no universal
principles can be found.
Procedural vs Declarative
The procedural - declarative controversy
revolves around the question
of knowledge as knowing how or knowing that.
The procedural approach assumes that
a person's knowledge of the
world is embodied in procedures that actively interpret the environment
and operate on it.
The declarative approach assumes
that knowledge is a collection of
facts that can be stated in logical propositions, conceptual graphs,
or other symbols.
@ Refer to the example
in the text (p. 5) about the condition played on Marvin Minsky, by his
graduate students.
# The notion of a mind's eye that observes
images in the brain: presumably the mind's eye would transmit stimulation
to a mind's brain, which would have its own mental image observed by another
mind's eye and so on in an infinite regress.
* Competence is an idealised knowledge
of language; Performance is the actual use of language in speaking and
understanding.