Chapter 3 Cognitive Science

homelybrrrInternet και Εφαρμογές Web

4 Δεκ 2013 (πριν από 3 χρόνια και 6 μήνες)

74 εμφανίσεις






Chapter 3


Cognitive Science


In this chapter we see how a renewed interest in the mind coming from various
disciplines has led to a joint endeavor.



Introduction


I
n the last few decades, a new science of the mind has emerged that goes much beyond
p
hilosophizing (although there is nothing wrong with that in itself). This new science is
called
Cognitive Science
(
CS
)
, which is usually described as the study of
(
mental
)

processes underlying intelligent behavior

(of humans,

animals, and even machines).
Scientists in this field go beyond philosophy in that they take into account findings in a
variety of fields and then construct explicit models of the mind (or a part of it), which
they often implement in the form of
computer
programs
. In this way they try to
simulate

the
hidden

processes

(especially of human minds)

to come to an understanding
of the
se processes and the

mind
as a whole
in explicit, scientific terms. In as far as
cognitive science is also interested in “intellig
ent machines,” there is a close link to the
field of
Artificial Intelligence

(
AI
)
, which literally attempts to construct programs
that
can make machines (‘
robots
’) see, hear, grasp and move.

In many cases
the goal of AI

is
to mimic
and model
human intelligence in a realistic way, that is, to develop models of
how the human mind actually works
. In
other cases, the goal is to develop machines that
can achieve specific tasks without necessarily claiming that these machines
(if
successful) “reason” in the way that humans do. In practice, most of cognitive science is
focused on intelligent behavior of biological entities (humans and animals), and for that
matter mostly on humans.

This difference between trying to understand ho
w the mind
drives behavior and trying to build machines that do things we need them to do is also
visible in the development of computer programs. One could imagine that we try to
model how a human being, being bilingual, might go about translating a text
from one
language into the other. We suspect that a human will analyze the sentence, grasp its
meaning and then express
es

the same meaning in that other language, using the rules of
that language. This is not necessarily how computer programs (like Google
translate) do
their job. Computer programs can use
large databases with whole sentences, or partial
sentences stored memory, each with the corresponding sentence in the other language, so
that translating is more like looking up a phrase in the list and th
en print out the
corresponding sentence in the other language that is stored with it. Computers are much
faster than the human brain in accessing large ‘memories’, so this technique works
reasonably well. The point here is that successful computer programs

do not necessar
il
y
serve good models for how the human mind works.

CS is a typical
interdisciplinary field
. It combines the efforts of all scientists
who work on (aspects of) the human or animal mind. According to some people we
should speak of the cognit
ive scienc
ES

(plural) because, to date, there is no unified
approach, and often various sciences contribute pieces of the puzzle more or less
independently

(
which sometimes lead to
miscommunications)
. As one example, I must
mention the complaint within CS
that linguistic theories of the mental grammar,
especially those that are inspired by Chomsky’s more recent ideas, are difficult to
understand or integrate into g
eneral models of the human mind that take into account how
the actual brain works. I will come

back to this point later on.

Within CS, a central science is, of course,
Cognitive

Psychology
, a branch of
psychology that has always been devoted to the study of mental processes, including
neuropsychology

(in particular the study of cognitive disorders
due to brain defects). In
addition, we find the contributions of many other areas of study:


-

Computer science

-

Artificial intelligence

-

Linguistics

-

Anthropology

-

Neuroscience

-

Philosophy

-

Ethology (the study of animal “behavior”)

-

Economics

-

Sociology

-

Political s
cience

-




This is not meant to be an exhaustive list. Also, most of the sciences mentioned have
specific subareas that are referred to as “cognitive,” such as cognitive anthropology,
cognitive linguistics, cognitive neuroscience, and the like. This makes s
ense because in
most cases, the relevant sciences have other concerns

as well which are less focused on
the human mind.

The fundamental assumption of cognitive science is that
cognition is information
processing
. This means t
hat the human mind is an
input
-
output system
. Information of
whatever kind is somehow perceived (input), and this information is stored in a memory
from where it can be called up
and produced in some way
(output).
B
efore information is
produced in some out
put
, this information can be processed or transformed (i.e.,
computations

of various kinds can be performed). This view entails the idea that stored

information takes the form of
symbolic representations

(i.e., expressions

consisting of
units that represent something
such as

innate information or
sensory information).



Origins of the Field


C
S finds its origin in the old philosophical debate concerning the nature of the human
mind (see chapter 2), but when we look at the r
ecent past we see a flurry of interest just
before, during, and after World War II (1939

194
5).
1

According to some, the field of CS arose in the late 1950s

some say during a
meeting in 1956 at MIT

which brought together researchers from various fields, all

sharing an interest in the working of the human mind. However, both Gardner and Dupuy
mention the fact
that this conference was preceded by a host of other
relevant
conferences, starting as early as 1946. In fact, between 1946 and 1953 there were as
many
as ten conferences (nine in New York and one in Princeton) known as the Macy
conferences (after the major sponsor), which brought together many of the leading
‘mind’
scientists around at that time. Gardner also refers to another conference (not part of the

Macy sequence) held in 1948 at the California Institute of Technology, where
, in part,

the
same people
got together
.

This whole period must have felt
to those who were part of this
‘cognitive wave’ as very exciting.

Prominent figures
(all extremely
smart

people with a diversity of interests)
were
Norbert Wiener
(
a mathematician;
1894

1964),

the mathematician
John von Neumann

(
1903
-
1957)
Claude Shannon

(
1916
-
2001;
the inventor of a science called Information
Theory
),
Warren McCull
och

(
1889
-
1969;
mathematician and neurophysiologist),
Walter Pitts

(1923
-
1969; a logician who developed idea about neural circuits),
Karl
Lashley

(
1890
-
1958;
a
behaviorist

psychologist

who realized that the cause of behavior
lies in the brain
), and
many
ot
hers. In addition, the Austrian
Ludwig von Bertalanffy

(1901

1972) is often seen as another important figure in the cognitive movement. He
developed ideas on a unification of different sciences (including biology, psychology,
etc.) in what he called Genera
l Systems Theory. All these researchers shared an interest
in the workings of the human mind.

Next, I single out
some key issues in the various
fields of research that these eminent scholars were involved in and
that proved to be
especially important for l
ater developments.



Information Theory


S
hannon focused on the notion of information processing, trying to measure the
information content of any given message

and also the amount of information that can be
reliable sent via som
e communication channel, given its limits. This lead to ideas about
data compression that proved very important in the development of computer storage.
Shannon also paid attention to
how information can be obscured by “noise” that
interferes with the signa
l. For this reason, he noted,
that information units (symbols) that
have a high degree of expectancy in a message can be short (compressed), while
unexpected units (that need to be robust to noise) need to be coded more extensively
(more redundantly). When

applied to language, we find the effect of these principles in
the fact that high frequency words (which therefore have a high rate of expectancy) are
usually short (‘the’, ‘a’, ‘me’, ‘in’ etc.) whereas words that are not so common are longer



1

Extensive overviews of these developments can be found in Gardner (1985) and Dupuy (2000).

(phoneme, un
ification, chapter, etc.). This also correlates with the fact that high
frequency words are often members of small closed classes (articles, pronouns,
prepositions), whereas words such as nouns, verbs and adjectives belong to large open
classes which make
the decision of which one
the hearer hears more difficult, which is
why such words must be longer, especially given that language is typically used in
circumstances where there is interfering noise.
Inspired by the results of neuroscience,
many looked at t
he brain’s neurological architecture, seeing neuronal communication as
analogous to information processing.



Cybernetics


N
orbert Wiener saw the brain/mind as a self
-
regulating system that, through a process
called
feedback
, could display goal
-
directed behavior. (Feedback involved checking the
“output” or “current state” of the system and comparing it to some intended goal,
allowing adjustment of the system to increase the match between the current output/state
and som
e intended output/state.) He compared the mind/brain to a thermostat that
controls temperature by comparing a desired state (the temperature set) to an actual state
(the temperature in a room) and making appropriate adjustments
when

there is a
mismatch bet
ween the two. He also came up with the name
cybernetics

for the general
study of “control (through feedback and adjustment) and information (processing)” from
the Greek word
kybernetes
, which means “steersman.”

V
on Neumann
, another important
figure in the
cybernetic movement

compared the mind to a computer program (an
analogy that we introduced in the previous chapter and that has stayed with Cognitive
Science until today).

The cybernetic movement was extremely energetic and, to many, very promising.
There
are, to be sure, important differences between the beliefs of the cybernetics group
and the beliefs of modern cognitive science. In particular,
despite the fact that cybernetics
aimed at the study of feedback
-
based systems at a very abstract and generalizi
ng level,
the cyberneticists believed that they could capture the essence of the human mind by
studying the neurophysiology of the brain (in that sense, th
ey were hardcore physicalists).
The relevant feedback system for them was in the workings of the brai
n. M
odern
cognitive scientists make a clear distinction between the workings of the physical brain
on the one hand and the notion of
cognitive representation

which is more oriented
toward analyzing the functionality of the
mind (cf. functionalism, discussed in the
previous chapter

and the discussion of the so
-
called tri
-
level hypothesis later in this
chapter
).

The terms cybernetics and the movement itself was superseded by approaches
labeled Artificial Intelligence and Cogni
tive Science, but that should not take away from
the fact that all these labels cover a common broad interest in the workings of ‘intelligent
systems’ that take input and produce regulated output/behavior and that learn and grow in
complexity.




Psycholo
gy


T
he emerging interest in studying the mind scientifically involved a departure from an
influential trend in psychology, called
behaviorism
, which had been dominant for
several decades (at least in the United States)
; see also chapte
r 2
.
J. B. Watson

(1878

1958) and
B. F. Skinner

(1904

1990) were psychologists who, building on earlier work
of the Russian psychologist
Ivan Pavlov

(1849
-
1936), argued that human behavior can
only be understood by studying

the observable behavior itself a
nd

the factors in the
immediate environment that apparently cause that behavior. They argued that there was
no point in spending time on attributing the causes of behavior to mental processes,
which, if present at all, cannot be studied scientifically anyw
ay. In the previous chapter, I
referred to this branch of behaviorism as methodological behaviorism. The behaviorists
were materialist in denying the mind or, to put it differently, in reducing what some call
the mind to observable external (and thus mater
ialistic)
events
. Thus, behaviorists like
Watson and Skinner were not willing to include the concept of a mind in their analysis of
behavior and learning. They believed that the only things you can see, as a researcher, are
the factors that are present in
the environment (things, natural or cultural events, other
people’s behavior) and what kind of behavior occurs. You can relate the observed
environmental factors (stimuli) to the observed behavior (response), and that is
all

you
can do. There is nothing in

between the stimuli and the responses that one can (or needs
to) talk about, and there certainly is no reason to believe that the response is determined
by anything but the stimuli.

With their denial of the mind, behaviorists therefore did not
(in fact, c
ould not) believe in a priori (innate) knowledge. As a result, behaviorists and in
particular Skinner, placed themselves squarely in the “nurture” camp by rejecting the idea
that human behavior is due to forms of innate knowledge. They were committed to th
is
viewpoint because they believed that human behavior cannot be accounted for in terms of
mental states or processes, which necessarily precludes any appeal to
innate

mental
processes.

Watson and Skinner argued famously that human learning takes place pur
ely on
the basis of environmental factors. By manipulating these factors (in terms of handing out
rewards, positive reinforcements, or punishment), organisms (humans and other animals
alike) can be made to do (or learn) anything. Skinner applied these idea
s to all forms of
learning and behavior
(doing many experiments with pigeons)
and devoted a whole book
called
On Verbal Behavior

to
the way children learn
language.

Behaviorism was attacked by Karl Lashley
(who studied with Watson and
practiced behaviorism

for many years)
during the 1948 conference mentioned earlier.
Lashley argued that language in particular was much too complex to be understood in
behavioristic terms. This position was supported by a young linguist named
Noam
Chomsky
, who wrote a lengthy
review of Skinner’s book about language

in 1959
,
arguing that his position with respect to language learning was untenable. According to
many, Chomsky’s review was a death blow to behaviorism, opening the door for a new
approach to language learning, one t
hat incorporated a serious study of the mental
processes underlying language behavior

At the same time, Chomsky suggested that
language learning must be based on innate abilities.



Computer Science


B
esides the direct attacks on
behavi
orism by Lashley and Chomsky, another
factor that
contributed to the idea that the mind should and could be studied fruitfully was the
invention of the
digital computer
. The
idea

of computers (in the form of abstract
algorithms) that can “solve” prob
lems in a finite number of

steps had arisen much earlier,

before World War II,
specifically
in the work of
Alan Turing

(1912
-
1954)
.
(In fact,
Charles Babbage
(1792

1871)
, long before that actually partially built a mechanical device,
called the Analytical

Engine, that could execute mathematical functions, and he
conceived of even more complex machines that he was unable to realize for lack of
money.)
However,
in more modern times,
in the late 1950s,
researchers at the University
of
Pennsylvania

(advised b
y
Von Neumann
)

actually built a device that worked in
accordance with Turing’s insights

(called ENIAC)
. The notion of such computing devices
(i.e., machines that would run programs) formed a new inspiration for thinking about the
human mind, in that it cou
ld perhaps be said that the relationship between the human
mind and the human brain was much as that between a program (software) and the
computer device (hardware). In short, it seemed attractive and suggestive to think of the
human mind as a kind of comp
uting program taking inputs, doing operations on these
inputs, and creating outputs.
This stimulated the idea of seeing the
mind as a
n
information
-
processing device, a device that takes input and produces output.



Cryptology


S
till othe
r factors may have contributed to the renewed interest in the study of the human
mind. During World War II tremendous efforts had been put into several areas that
involve
cryptology
, the art of coding information in such a way that it could only be read
by

people possessing the key to decipher the code. These efforts led scientists like
Alan
Turing
to develop general theories of (human) problem solving and information
processing
.

During World War II, Turing and many other mathematicians and scientists
were
recruited by the British government to help in deciphering of the various codes that
were used by the Germans
.



Neuroscience

(Neurology, Neurophysiology, Neuropsychology)


A
nother way in which the war contributed to the study of the h
uman mind was by
causing many war casualties, in particular people with all sorts of brain damage and
resulting deviant behavior. Neuropsychologists, such as the famous Russian
Alexander
Luria

(1902
-
1977)
, came to understand
that
the brain
/mind
, is a cruci
al determinant of
human behavior. These realizations further undermined the behaviorists’ dismissal of the
mind and their claim that behavior was fully dependent on stimuli in the environment.

In the United States, the emerging field of neuroscience had gr
own considerably
and, in the 1960s, it was felt that the time was ripe to link the knowledge about the brain
to the various ideas and theories about this thing called “mind”. This feeling, that the gap
between the brain and the mind needed to be bridged, w
as an important incentive, along
with the other factors that have been mentioned, for the development of a new field,
cognitive science.



Symbolic representations


A
n important ingredient of the Cognitive Science enterpr
ise is that notion of
s
ymbolic
representations

(also often called
cognitive representations
)
.
Symbolic representations

consist of abstract cognitive units

organized in some (linear and/or hierarchical) way
.
These units that have some sort of interpretation

(or meaning) associated with them,
which is why they are often called
symbolic
, since a
symbol

(in the
semiotic sense
) is a
package of a meaning or referent and a form. (For ease of grasping this idea, think of a
symbolic representation a
s a formula on paper consisting of discrete graphic units such as
letters, numbers and ‘>’, ‘<’, “=”, etc. In other words, a symbolic representation is
something like a mathematical expression of some sort.)
There is no necessity to be able
to identify a p
hysical basis for such symbols (i.e. we leave open whether cognitive
symbols correlate with specific identifiable neural structures).
When we write, in fact, we
create a symbolic representation in which the letters refer to the sounds of language.
Likewise
, we can say that in the mind we have a symbolic representation of the sounds of
language as well, consisting of discrete units that stand for the sounds.


A question of some importance is whether the symbols that make up the symbolic representation stand
in a
‘natural’ relationship to the mind
-
external realities that the
y
are supposed to represent. If we think of these
symbols as being somehow re
sembling what they represent then
these symbols are actually
icons
, since in
Semiotics, symbols are taken to sta
nd in an arbitrary relationship to what they represent.


Cybernetics did not postulate a symbolic level of representations as distinct from the
physical level (at least not consistently), even though the cyberneticist Von Neumann,
who compared the mind/bra
in distinction to the program/computer distinction, inspired,
by this comparison, the functional perspective that uses symbolic representations.

Mental operations (
computations
) make changes on
cognitive, symbolic
representations by performing addition or
deletion or permutation, etc.
This introduces a
distinction between an input, or basic level of representations

which is derived from some
external source (for example perception)
, and an output, a derived level. Cognitive
science creates models of the min
d in terms of representations and computations. It is
believed that units and operations somehow correspond to actual, physical brain
activities, but many cognitive scientists believe that it is not crucial for the cognitive
enterprise to demonstrate what
these brain activities are or where in the brain they
happen.
We saw in the previous chapter that this ‘neglect’ for the physical basis is
associated with functionalism. Many c
ognitive scientists
take this functionalist stance and
thus
believe that it is u
seful to talk about the mind abstractly, that is, in terms of how it
functions, without necessarily knowing at the same time how the brain implements the
representations and operations. It is fair to say, as we did, that the cyberneticists, even
though the
y saw the distinction between talking about the workings of the brain
schematically
(i.e. functionally)
and in physical terms, assumed a much closer connection
between the two than we see in cognitive science today. As a consequence, as Dupuy
claims, the c
yberneticists
were not inclined to
view their models as involving symbolic
representations. They had a much more mechanical view of their models, almost as if
they were representing the technical structure of some machine.
In the next section, we
will see
that these various approaches are not incompatible.



The Tri
-
Level hypothesis


A

characteristic of modern cognitive science is to
say

that we can study cognition at
three different levels. This distinction is due originally to

David Marr

(1945

1980), a
pioneer in cognitive science and the study of visual perception. In the simplest terms the
three
-
way distinction is easy to state; in practice, cognitive scientists may differ in
where
the
y

exactly draw the line between these lev
els
.



The
functional
Level


I
n
doing

cognitive science, we start out by being interested in some
question

regarding
(let us say, human) cognition. The phenomenon could regard the visual system and a
specific question could be:

How does the mind manage to transform the two
-
dimensional
image that light projects on the retina of the eye into a three
-
dimensional mental image?
Or, a linguist could ask: How does the mind transform the language input that a child is
exposed to into a
mental grammar of the relevant language? Or: someone might ask: How
do people (learn to) play chess? In all cases the issue at hand is called “a problem” that
needs to be solved by a particular part of the mind, and the central task of the cognitive
scient
ist is to come up with a model of the solution, that is, of the way that the mind does
it.

At what is called the
functional
level of analysis we state and
analyze the
problem

that we wish to focus on.
We make a
problem analysis
.
This goes beyond
posing a m
ere question. It is one thing to ask: How do children acquire their mental
grammars? But before we proceed with specifying the solution, we need to characterize
the problem
in much greater detail
. This means that we have to analyze in precise terms
what th
e
nature is of the input and output
. In the case of linguistics, this means that we
have to specify the kind of language input that children are exposed to
. We also need to
specify the nature of the knowledge that the child needs to acquire (i.e. the natur
e of the
mental grammar) in order to produce correct, grammatical sentences (such as
John
ment
ioned that Sue saw that she fell
).
In short
,

at the
functional

level we state
what the
cognitive module is supposed to do, what its function is.


A
t this most gen
eral level of analysis, it may appear that the problem that we are
trying to tackle is too big and needs to be split up into a number of smaller,
interconnected problems. This conclusion is typical in linguistics because when focusing
on language, we quick
ly realize that language is very complex, which means that learning
language poses not one, but several
smaller
problems. For example, figuring out the
sound system may be a rather different problem than figuring out the sentence structure.
Thus the proble
m analysis
may lead to specifying various submodules and the
relationships between them.

The Algorithmic Level


H
aving come to proper understanding of the problem, we can now proceed and actually
construct a model of the cognitiv
e module that relates input to output.
Where we thus far
had a rough sketch, we now proceed to flesh out the details.
The crucial question
s are
now
:
what is the nature of the symbolic representations that represent the input (which
symbols are we using and

what do they stand for). Also:
Which information processing
steps constitute the procedure that leads from input to output?
What are the constraints on
information processing?
A

technical term for a step wise procedure
is

algorithm
.
Natu
rally, we are not just interested in any algorithm that does the job; we wish to come
up with one that models closely how humans actually do the job. We want the algorithm
to be
psychologically realistic
.

When we have written the algorithm (almost like a c
omputer program) we could
say that we have constructed an explicit
functional model

of some cognitive capacity
. So
far we have said
nothing about the physical brain mechanisms that are involved. We
worry about that at a third level

of analysis.



The Implementational Level


A
t this third level we ask: What is the nature of the physical device that runs the
program? In the case of humans, this device is the brain. Thus, we need to know how the
brain wor
ks, and this means studying neurons, neural firing, neural connections, and so
on. According to some cognitive scientists the implementational level is the most
concrete level of analysis, and is indispensable, in particular if a physicalist interpretation

of the mind is adhered to. After all, if the mind results from processes in the brain, it
would seem obvious that a model of a certain cognitive capacity, notably the algorithmic
analysis, should reflect or be compatible with what is known about the worki
ngs of the
brain.



Why the tri
-
level hypothesis is good for cognitive science


In the previous chapter we say that a functionalism proclaim that the imple
m
entational
level is not relevant for understanding how the mind works. In this chapter we have seen
that cyberneticists precisely focused on brain mechanism in their attempt to understand
how intelligent systems work. If we accept the tri
-
level hypothesis we make room for
various approaches, without claiming that one is more important than the other. Thi
s
leaves room for what is
often called the reductionist view,
which is
that ultimately the
functional/
computational level and the algorithmic level can be derived from a complete
analysis at the implementational level. To go back to our discussion of the m
ind

body
problem in chapter 2, we might say that such a reductionist view is in line with hard
-
core
physicalism.

It would seem that the cyberneticists were not that extreme. In any event, it
would seem that we are a long way from understanding the mind in
purely physical terms
and it may be that the hope that neuroscience will one day provide all the answers is
misguided

because at the end of the day precise knowledge of brain processes cannot
explain why these and not other processes occur and what they ar
e good for.



Additional Levels of Analysis


T
he
three

levels,
however,
do not exhaust what can or needs to be said about cognitive
abilities. I will mention at least three other levels, namely the behavioral level, the genetic
level, and the evolutionary
level.



The Behavioral Level


C
ognitive models
often stop short of giving an account of
actual behavior
, but it is
crucial to see that there is a distance (a gap to bridge) between the
cognitive

analysis
(even if it covers all thr
ee previous levels)
and the actual behavior.
This is because the
input string and output string (that are mediated by computation) are cognitive, symbolic
representations
. Taking language as an example, we also need to specify the relationship
between the
input string and the auditory system. And on the other side, we need to
specify the relationship between the output string and speech production. What I mean
here by behavior is thus the perception and production system of language. (It is perhaps
odd to c
all the act of hearing ‘behavior’, unlike the act of speaking which feels like
actually doing something. For want of a better term, I will use the term behavior for
both.)



Sound wave














Ear




A: PERCEPTION


Nerve signal to the brain creating










input representation














computations



B: COGNITION













output repr
esentation












Nerve signal to articulatory systems









C: PRODUCTION





lungs/jaw/tongue/larynx










Sound wave



The three level
s

discussed before are approaches to box B (the cogn
itive system, in this
case the language module or mental grammar). Clearly, a full analysis of language would
have to involve models of the perception and production systems as well.
This is what I
here call the behavioral level.




Some terminological iss
ues


-

Usually, the first level is called the computational level. Here I use the term
functional level instead. This is because the term ‘computational’ seems more
applicable to the algorithmic level, which indeed is confirmed by people who speak
of the alg
orithms that relate input to output as ‘computations’, a terminological choice
that I have followed here.


-

Recall that, according to the tri
-
level hypothesis
,

implementation

refers to the neural
activity that corresponds to the input, computation and outpu
t of the cognitive
module. The neural basis of output representations (or those part
s

of the output that
bear on the spoken aspect of language) could be neural connections in the motor
cortex that
stimulate the nerves that

activate
the articulatory organs
for speech.


-

Confusingly, linguists sometimes use the term ‘implementation’ for relating the
output
r
epresentation to actual speech; they use the term ‘phonetic implementation’
for this.

It
is more adequate to see
the relation between
the cognitive repres
entation of
speech

and actual
speech

events is one of
interpretation
, i.e., of “meaning,” because
it might be said that the symbols that make up the
output

representation are
interpreted
as, or stand for

these phonetic events.


-

As we will see below, Choms
ky makes a distinction between
competence

(which is
the cognitive system) and
performance

(which is production of language). This
shows a bias to the production part of language behavior, unless we take performance
to include perception.



The Genetic Leve
l


I
f we take the view that the cognitive modules that we are analyzing contain structures
and information that are innately present, we are committed to raising the question of
how the innate specifications are anchored in the genome
. In other words, we need to
analyze the cognitive faculty at the genetic level. With respect to most cognitive faculties
this kind of analysis is still in its infancy. With reference to language, we will discuss
some of its supposed genetic
grounding

in c
hapter 18.






The Evolutionary Level


A

further level, which is relevant only to modules that are (partly) innate, is to look into
the evolutionary processes that have led to the genetic specifications of these modules.
With r
eference to language, we will address these issues in Part VII of this book.




The Cognitive Science Paradox


L
et us mention that there is an interesting paradox when it comes to modeling human
intelligence. So far it would seem
that
problems that humans
find rather easy (such as
acquiring language) are much more difficult to model than problems that are much more
difficult for humans (such as playing chess). All people (barring medical conditions) are
“master” speakers of at least one language, whereas ve
ry few people are master chess
players! Let us call this the
Cognitive Science Paradox
.
It has been argued that things
that come easy to people most likely are easy because we do them on the basis of innate
knowledge. This

innate knowledge is likely to be quite complex in some cases. In
contrast, things that are hard to learn or do might be hard because we have no innate
‘help’ which means that we have to build up the relevant knowledge from scratch.



Two Different Approac
hes: The Classical and the Connectionist Architecture


W
e have seen that
according

to
some cognitive scient
i
sts

one cannot develop a
cognitive

model (that addresses the first two levels) in complete isolation from what is known
about the workings of the br
ain, especially not if one takes a physicalist stance (which
says that the mind results from or is in some sense equal to the workings of the brain). In
other words, whereas the first two levels of the tri
-
level hypothesis seem to deliver a
functional
mode
l of what we might call the modules of the mind, or mental capacities, the
implementational level is required to anchor this model to a model of brain architecture
and brain processes.
Taken to an extreme, this might lead to a view which finds the
developm
ent of a functional model including algorithms a waste of time, especially if it
seems that the very nature of an algorithmic approach is in conflict with what is known
about how the brain works.




Classical Architecture
:

Se
rialism


O
ne approach to the task of modeling an intelligent system (like the human mind) is
based on the idea or belief that information processing is a
serial

step
-
by
-
step and thus
algorithmic
process. The most original conception of in
formation processing as a serial
algorithmic
process is due to the English mathematician
Alan Turing

(1912

1954). In an
attempt to show that certain mathematical problems can be solved in a finite number of
steps, he developed a problem
-
solving “machine,”
which was claimed to have the formal
power to solve any (solvable) problem in a finite number of steps. This machine is now
called the “Turing machine
.” This machine can be envisaged as a typewriter that can read
and (over)write symb
ols on a tape. In doing so, it can move back and forth along the tape.
Here, we will not go into a detailed description of this machine, but simple as it sounds,
this idea provided one of the bases for our modern digital computer
s. (
The Turing
machine mode
l is a serial model. A “second
-
order” Turing machine is called the
Universal Turing Machine
, which can compute everything that any specific Turing
machine can compute. This is also still a serial model.
)
Turing machines are

not actual
physical machines; they are pencil
-
and
-
paper algorithms and as such not dependent on
any specific implementation.
The development of the digital computer has shown that the
basic idea of the Turing machine can be implemented in the form of actu
al devices.

The question now arises whether the brain works like a Turing machine. From the
start it was quite obvious that this was not the case, at least not in all respects. First,
Turing machines are cumbersome and slow. If you think of the tape as the

memory of a
system, a Turing machine could only move through this tape (memory) in a strict linear
fashion; to go, let us say, to a spot near the beginning of the tape, the machine would
have to move through all intermediate positions and then back throug
h all these positions
to continue and so on. Human memory doesn’t work like that. Rather, we have what is
called
random access
; we can, apparently, jump from one memory cell to the other. To
get rid of this problem, cognitive scientists

conceived of varian
ts of the Turing model that
would allow random access of memory cells. One such model was called a
physical
symbol system

(Newell 1980). (The idea to invoke random access was first suggested by
von Neumann; hence we also spea
k of a von Neumann machine.) Putting details aside,
this kind of system has a random access memory and can thus operate much more
quickly. Crucially, this type of system is not more powerful (i.e., it can’t solve more
problems) than a (Universal) Turing ma
chine, and it is still serial in that it can only
perform one operation at a time. (Despite the term
physical
, such systems are still paper
-
and
-
pencil models that could in principle be implemented in any type of physical
machine.)


What we have just discus
sed is called the Classical Architecture of Cognitive
Science. It has been pointed out that there are three kinds of problems with this
architecture, given that it is meant to model the human brain/mind:


-

The serial nature of information processes makes it

very
slow
. Assuming that certain
algorithms might involve thousands or even millions of steps to solve a given
problem, a problem that humans solve rather quickly, serial processing seems
unrealistic. Here we have to take into account that the neuronal co
unterpart of one
computing step would be something like a neuronal firing. We know that the rate of
neuronal firing is at maximum about 200 times per second. This may seem fast, but if
the solution of a given problem that humans solve in a split second tak
es thousands of
steps, it would take our Turing machine a minute or more. (Modern computers that
are serial don’t have this problem because we now have chips that allow several
millions of steps per second. However, this by no means implies that computers
with
a serial architecture are fast enough to do what humans do.)

-

The serial nature makes information processing
vulnerable

to small defects. Any
chain is only as strong as the weakest link. Serial processing entails that if a certain
symbol is misread, or

a certain
calculation

is misapplied, the whole process is likely
to crash. Again, it does not seem that human intelligence is so vulnerable. Although
crashing is possible, more often than not an alternative route is found that leads to the
required soluti
on.

-

Digital processing is strictly based on “yes” or “no.” A symbol on the tape either is or
is not the symbol that triggers a certain operation. There is no in
-
between, no third
option,
no gradience
. Processes that require a gradient approach are called
a
nalog

(as
opposed to
digital
). Turing machines and modern computers are digital, and this
makes them
inflexible
. In real life we seem to be dealing with gray areas all the time.
Things are not black or white; some things are better than others on a scale o
f some
sort.


To deal with these problems, a new brand of cognitive scientists proposed a different type
of model, called the connectionist model.



Connectionism
: Parallelism



C
onnectionism

is based on a different architecture of human cognition, namely a
parallel architecture
, which is claimed to be closer to how the brain actually works. In a
parallel architecture the information flow is not modeled as a single chain of steps, but
rather a
s a multitude of such chains that happen side by side. Why is this more realistic?
Consider what is known about the workings of the brain. Information processing at the
neuronal level takes place when a neuron “fires.” This means that it generates an inter
nal
electrical current, which leads to the release of chemicals into a small space called a
synapse (which is essentially the fluid
-
filled area between neurons). These chemicals are
“sensed” by other neurons, which in their turn generate an internal curren
t (or, if the
neural transmitters are inhibiters, prevents that process). Crucially
,

the chemical produced
by a given neuron is sensed by many other neurons (up to thousands). Thus information
travels in a parallel and distributed way through the neuronal
network. Accordingly, this
kind of processing is called
Parallel Distributed Processing

(PDP)
.

A PDP model can be represented by a diagram made up of two kinds of entities:
units

(also sometimes called nodes, a
nalogous to neurons) and
connections

(represented
by
lines,
modeling synapses). On one end of the model, we find
input units
; on the other
end we find
output units
. In between these, in a model o
f any sophistication, we find
hidden units
. Thus, we get a multilayered system that produces a certain response given
some stimulus input:





o

o

o


Output units






o

o

o

o

o


Hidden

units





o

o

o

o

o

o

o


Input units


This model suggests that networks are hierarchical, which is stimulated by certain results
in the neuronal processing of visual information by the Nobel Prize winners
David
Hubel

and
Torsten Wiesel
.

Being analogous to

neurons,
hidden
units process information in three steps. First,
they combine the effect of all incoming signals. Secondly, the unit computes its level of
activation (which is based on the effect of the incoming signals), and third, it produces its
own si
gnal (based on its activation level), which is then sent to other units at the next
level. The flow of information can be in one direction only (input to output) or allowed to
go in both directions.

The connections (being analogous to synapses) can amplify

or inhibit the signal
being sent; this is indicated by saying that these connections have a certain
weight
. The
signal being sent by a unit is multiplied by the (positive or negative) weight value of the
connection. Thus, the connection wei
ght indicates the strength of the connection. The
entire set of connection weights is called the network’s
pattern of connectivity
,

and it is
this pattern that determines how the information flows.

A noteworthy property of n
etworks is that they are not “programmed” (i.e.,
provided with a pattern of connectivity) in advance to do a certain job. Rather, the idea is
that networks
can
“learn” how to do the job, and thus it is said that the network must be
“trained.” In a training

session, one presents a certain input to the network for which the
response is known; all the weights
of the connections
are set at arbitrary values. The
response produced by the network is likely to be wrong, which we can see by comparing
it to the corre
ct response that we already know. We then adjust the weights and try again.
A certain procedure, which I will not discuss here, will tell us how the weights can be
adjusted so that a better result is obtained
each

time.
The procedure optimizes the
function
ing of the networks.
After a number of trials, the weights of the connections end
up being such that the output response no longer contains errors. We now say that the
network has “learned” a stimulus
-
response sequence.

Proponents of connectionism claim th
at this way of modeling the mind overcomes
the problems that were mentioned earlier with the classical serial architecture. Briefly,
parallel processing leads to speed (many things happen at the same time), it is more
robust (if one connection path breaks
down, the others still work), and it can produce
gradient solutions. On top of all this, connectionists claim that a network model presents
a more realistic picture of how the brain actually works. In other words, a network model
(which we locate at the co
mputational and algorithmic levels) squares smoothly with an
analysis at the implementational level.

But there are also problems with the connectionist architecture. Even though a
network may be successful after training in finding a solution for whatever
problem it is
designed for (e.g., discriminating between two patterns), it is actually unclear how the
network does this (i.e., by looking at a collection of weights that seem to do the job, a
researcher might not be able to tell what the generalization re
ally is). We know that there
is an algorithm in there (because the system seems to work), but the algorithm is scattered
over the network’s pattern of connectivity.




Who Wins?


A
t present the debate between classical and connectionist cognitive scientist
s is still very
much alive. There is more at stake here than the proposed architectures. Even though
there are many varieties of network designs, connectionists generally agree that
the
network approach can be seen as a characterization of the human mind a
s a general
learning device. The network architecture is not domain
-
specific. The same network
approach can be used to learn whatever needs to be learned.
Also, networks do not have
built
-
in knowledge.
They start from scratch.
Even though they start out wi
th a certain
pattern of connectivity, this pattern is random and only becomes specific in the course of
training. Classical cognitivists on the other hand usually assume that their models are
domain

specific. Different
cognitive systems

require different a
rchitectures and
algorithms that the researcher has to come up with to construct the model. In addition,
classical models often have built
-
in knowledge. What all this entails is that when it comes
to the nature/nurture debate, classical cognitivists often
side with the rationalist view that
a cognitive system is (besides being
domain

specific) provided with a priori, or innate
knowledge, whereas connectionists more often side with the empiricist view that all
knowledge is obtained on the basis of
one type o
f general learning procedure
. It is not a
coincidence indeed that connectionist networks are similar to so
-
called classical
associationism (cf. chapter 2), which is the term for the empiricists view on learning (i.e.,
by association, analogy, etc.).



Fiel
ds of Study and Methodology


C
ognitive science is a broad field of study. In its broadest sense it tries to account for
the
way in which intelligent systems (a) obtain, (b) represent, (c) transform information and
(d) use information
to behave in a certain

way. This implies that CS has a special interest
in (a) input systems (i.e., sensory
-
based perceptual systems such as vision, hearing, etc.),
(b) memory systems (storage/retrieval of information in terms of images and language),
(c)
thinking,

reasoning, a
nd problem solving and (d)
output systems (language production
and other patterns of behavior, involving matters of motor control, and so on).

Cognitive scientists will often use experimental methods (using subjects and
instruments of various sorts) to inf
er information that can lea
d them to constructing their
theories
. They will also often try to test these theories using computer programs that
simulate

intelligent behavior
.

The idea is that if you claim to have an explicit model of
some cognitive module t
hat humans seem to have you should be able to write a program
and design hardware that simulates whatever it is that humans do.



Conclusions


W
e have now learned that there is an emerging interdisciplinary science of the mind.
Here, scientists working in
many different fields compare notes and formulate new and
detailed questions

and models
. Modern linguistics falls squarely within the goals of
cognitive science; it can even be said that the approach to language that originates with
the work of Noam Chomsk
y in the 1950s has been a crucial impetus to the emergence
and development of this field. It is therefore troubling that, as I mentioned earlier, people
in cognitive science fields
other
than linguistics often have problems keeping up with the
work of Chom
sky and his followers.
True, the tri
-
level hypothesis allows for different
scientists being focused on different aspect of a cognitive domain.
We thus need to
recognize that work in all sectors of cognitive science is specialized and, despite being
offered

in the collaborative, interdisciplinary spirit of cognitive science, not always easy
to penetrate for researchers who come from disciplines other than the one that offers the
results.
But this should not lead to a situation where different scientists are
totally
unable
to communicate
or even ridicule other people’s work
.
Fortunately t
here are many
linguists who, while still operating within the original and defining perimeters of
Chomsky’s approach, produce work that does not suffer from the same problems.