We look at cognitive psychology and
its relationships to linguistics and artificial
intelligence.
Assumption:
The brain interprets input from
the sense organs by assembling a model
of the environment. Thinking, talking, and problem solving are then based
on that model.
2.2.
Percepts
During perception, the brain keeps
a temporary record of the sensory input.
Neisser (1967) called that record an icon.
For a person to "see" a complete
figure or scene, perception must construct
a complete model out of many incomplete partial view.
Immanual Kant (1781) , Otto Selz
(1913, 1922) and Bartlett (1932) proposed
the notion of schema, which acts as a blueprint for a mental model,
to explain how perceptual mechanisms can correctly assemble a partial
view.
With the right schema, separate icons
are integrated into a stable image.
A schema is a pattern for assembling
units called icons.
Percepts are like prefabricated building
blocks derived from previous experience
and used to build models for interpreting new experience.
How people interpret sensory input
depends on their stock of percepts.
Hearing and touch also rely on percepts
and icons (i.e., auditory icon and
kinesthetic icon).
Apparently, there is no olfactory
icons and percepts for the sense of smell.
Sounds of language are interpreted
in called phonemes, which
corresponds to vowels and consonants.
Phonemes form syllables, syllables
form words, and words form phrases and sentences.
Question: Is perception bottom-up
or top-down process?
Some evidence favours a top-down
approach, and other evidence favours
bottom-up approach.
Psychological conclusion: both approachs
are valid and compliment one another.
This forms the basic principle of
AI:
top-down reasoning is a goal-directed
process that imposes a tightly
controlled organisation;
bottom-up reasoning is a data-directed
or stimulus-directed process
that leads to more diffuse chains of associations.
This two approaches may be combined
in bi-directional reasoning, which
is originally triggered by some stimulus in the data, but which
then invoke a high-level goal that
controls the rest of the process.
2.3.
Mechanisms of Perception
When the brain receives a new sensory
icon, it must search its stock of percepts
to find ones that match parts of the icon.
The search mechanism, called the
associative comparator, must have the
following characteristics:
Associative Retrieval - An ordinary
computer retrieves data by an
address in storage. The brain has an associative mechanism, which
retrieves the pattern that matches best.
Top-Down Match - Perception finds
percepts that match the overall pattern of an icon before it fills in for
the details.
Stimulus Constancy - Stimuli from
the same external object are recognised
as equivalent despite varying size, brightness, and retinal
position.
Distributed Storage - A particular
memory is not located at a specific
point in the brain. Lashley (1950) showed that an area of the
cortex can be destroyed without erasing the memory.
Perception requires other mechanism
beside the associative comparator:
visual buffer - on which images are
rotated, projected, and combined
with other images.
assembler - assembles and transforms
percepts, each of which matches
part of a sensory icon.
motor mechanism - organise parts
of image into a complete form.
In perception, the assembler generates
a working model that matches incoming
sensory icons.
The associative comparator searches
for available percepts that match all
or part of an incoming sensory icon. Attention determines which
parts of a sensory icon are matched
first or which classes of percepts are
searched.
The assembler combines percepts from
long-term memory under the guidance
of schemata. The result is a working model that matches the sensory
icons. Larger percepts assembled from smaller ones are added to
the stock of and become available for future matching by the
associative comparator.
Motor mechanism help the assembler
to construct a working model, and
they, in turn, are directed by a working model that represents the
goal to be achieved.
2.4.
Conceptual Encoding
A/S Ratio (A - Association Cortex;
S - Sensory Cortex)! Hebb
(1949) found that a high A/S ratio suggests a high potential for
sophisticated, intelligent behaviour.
Connections in the association area
develop from sensory input, thus, animals
with a high A/S ratio require a great deal of input to reach their
full potential.
A quantitative increase in the A/S
ratio can lead to a qualitative difference
in the complexity of behaviour.
Four mechanisms have been considered
for encoding information in the
association cortex:
Synesthesia - input to one
primary zone, such as hearing,
may directly stimulate an image in
another primary zone, such as vision.
Mental images - people differ
widely in how vividly they experience
images.
Language - the most detailed
encoding for external communication
is language.
Concepts - more abstract than
language are concepts and
conceptual relations.
Concepts are so abstract, thus, evidence
for them must be obtained indirectly.
This is much evident in the study of abstract thinking, when one
analyses how mathematicians develop mathematical ideas, and how
deaf children are better at abstract thinking compared with hearing
children.
Note that the ability to think abstractly
can develop independently of language
and scholastic achievements.
Language and logic are independent
skills.
To deal with language and imagery,
concepts must be associated with both
words and percepts@
.
Concepts may be associated with images,
but they are more abstract than
images.
When a person sees a cat sitting
on a mat, perception maps the image into
a conceptual graph.
A person who is bilingual in French
and English may say, in speaking French,
Je vois un chat assis sur une natte. In describing the same
perception in English, the person
may say I see a cat sitting on a mat.
The same conceptual graph, which
originates in a perceptual process, may
be mapped to either language.
Conceptual graphs are universal,
language-independent deep structure.
In AI, the term concept is used for
the nodes that encode information in networks
or graphs: a concept is a basic unit for representing knowledge.
Defining concepts as a unit presupposes
that concepts are discrete.
This assumption is supported by the
fact that discrete relationships are remembered
more accurately than continuous quantities.
Even if people cannot remember continuous
quantities, they can still detect
them. They cannot, however, encode them in long-term memory.
To adapt the discrete words to a
continuous world, natural languages have
"fuzzy" words like somewhat, very, almost, rather, more or less,approximately, just about, and
not quite.
Such words cannot provide a continuous
range of variability.
Zadeh (1974) developed a theory of
fuzzy logic to assign precise values to
such terms, but his calculus of fuzzy values makes distinctions that
no natural language ever represents.
Advocates of AI, who concentrate
on the discrete aspects, are optimistic
about the prospects for simulating intelligence on a digital computer.
Critics who concentrate on the continuous
forms maintain that simulation
of intelligence by digital means is impossible.
Since our brain use both kinds of
processes#
, a complete simulation may
require some combination of digital and analog means.
2.5.
Schemata
Concepts and percepts are building
blocks for constructing mental models.
Rules or patterns are required to
organise the building blocks into larger structures.
Kant (1781) introduced the term schema
for a rule that organises perceptions
into a unitary whole.
Selz (1913, 1922) used schema as
a basis for his theory of schematic anticipations.
Bartlett (1932) made the observation
that a schema is an active organisation,
and a schema must be operating in all orderly behaviour.
All complex behaviour shows the need
for schemata that organise elementary
units into larger patterns.
Read pp. 42 - 51.
In AI, Minsky (1975) showed the importance
of schemata, which he called
frames.
2.6.
Working Registers
James (1890) distinguished two types
of memory:
primary or short-term memory,
which maintains consciousness of
the immediate past;
secondary or long-term
memory, which is "knowledge of a former
state of mind after it has already once dropped form consciousness".
Although consciousness is a private
experience, it is correlated with measurable
activity in the cerebral cortex. Ref to experimental evidence
on pp. 51. Also refer to discussion of the functional difference
of short-term and long-term memory
(pp. 52 - 53).
Short-term memory does not contain
actual data, but pointers to previously
stored memories. In other words, short-term memory consists of a limited
number of working registers, each of which excites or
activates some record in long-term storage.
Miller (1956) showed that short-term
memory can hold about seven chunks
of information, where a chunk is the amount of information in a
schema.
Broadbent (1975) argued that a better
estimate is three working registers
rather than seven. Ref to his reasons in pp. 53.
Marcus (1980), in his PARSIFAL program,
further provided evidence that
three lookahead buffers were sufficient for a deterministic parser
if each buffer could hold an arbitrarily
large chunk.
2.7.
Recognition and Recall
Recognition memory is more accurate
than unaided recall.
There are two theories:
Threshold theory: a weak memory trace
is sufficient for recognition,
but a stronger trace, exceeding some minimum threshold,
is necessary for recall.
Two-Process theory: recognition is
a process for checking the familiarity
of an image, but recall involves a separate process of retrieval
or reconstruction.
Ref to discussion on pp 56 - 57.
The associate comparator$
and the assembler, which was originally proposed
as mechanisms of perception, can also serve as mechanism for
memory.
In recognising a sensory icon, the
associative comparator compares it with
all the records stored in long-term memory. For recall, the assembler
can join schemata to construct a memory probe. Then the associative
comparator can test the probe for recognition and retrieve associated
material for recall.
2.8.
Central Controller
Like instructions in a computer,
conceptual graphs are static data. To control
the linear flow of speech and other behaviour, some unit must convert
the static data into an ordered sequence of activity. That unit is
called the central controller.
Ref to evidence given on pg. 59 -
60 which point to the frontal lobes as a
major control unit.
Summary
All the components were proposed
by psychologists on the basis of traditional
psychological evidence. They fit together to form a system that
looks remarkably like an AI program:
The first step in an ordered chain
of thought is the selection of a conceptual
graph that anticipates the form of the desired goal.
Certain concepts in the graph are
flagged with control marks. Each control marks triggers expectancy waves,
which stimulate the associative
comparator to find matching schemata.
When the associative comparator finds
a matching schema, the assembler
joins it to the working graph. If the resulting graph satisfies
the control marks, it attains a state of closure, and the expectancy
waves are extinguished.
The result of joining a scheme to
the working graph may cause control
marks to be propagated to new nodes in the graph. The control
marks on the new nodes then trigger further searching.
The limited number of working registers
limits the number of control
marks that can be active at the same time. If there are more
than 3 unsatisfied control marks, earlier ones are suspended
until the more recent ones are satisfied.
When control marks for recent subgoals
attain closure, earlier control
marks are reactivated until the original goal is satisfied.
# Motion of tongue, lips and other parts
of the body must be continuous according to laws of physics. Sensory
channels for reci\eiving input are either continuous or discretised with
sch a fine
mesh that distinct units are not preceptible. Emotions are internal
states that depend on the continous
variability of hormones and other substances in the blood.
$ originally proposed as a mechanism
of perception