brains do, and the implementation at neural level is usu-
ally of little concern. Arguably, these theories are all
about psychological phenomena and the physical brain
should not even be considered in the context of this ap-
proach. Bennett and Hacker (2003), for example, believe
that
[. . . ] it makes no sense to ascribe such psychological
functions [i.e. perceiving and thinking] to anything
less than the animal as a whole. It is the animal that
perceives, not parts of its brain, and it is human
beings who think and reason, not their brains. The
brain and its activities make it possible for us - not
for it - to perceive and think, to feel emotions, and
to form and pursue projects (Bennett and Hacker,
2003, 3).
Essentially, the top-down 2 approach deals with high
level cognitive functions, and the brain, or entire being,
is viewed as a single black box, or a collection of black
boxes with certain functional properties.
Input nodes
Context nodes
Output nodes
Figure 2. SRN architecture
The bottom-up approach, in contrast, deals with the base
elements, namely neurons, and their physiological and
functional properties and processes. Functional aspects
of brains, or parts of brains, are investigated by look-
ing at individual neurons and structures of groups and
networks of neurons. Cognitive Neuroscience and some
work in Artificial Intelligence is concerned with how cog-
nitive functions might be implemented in brains.
Currently, methods are explored to connect the top-down
and the bottom-up approaches in attempts to ground
high-level psychological phenomena in neuro-physiology.
One such research program concerns the mapping or lo-
calizing of cognitive functions in the brain. Modern tech-
nologies such as PET and fMRI3 are commonly used for
2I am using the terms top-down and bottom-up in favor of
high-level and low-level to suggest that these are not static
research programs, but that they are dynamic endeavors aim-
ing to close the divide between them.
3Positron Emission Tomography (PET) and functional
Magnetic Resonance Imaging (fMRI) are based on the as-
sumption that mental activity causes an increase in the
metabolic rate of neurons, and therefore an increase in the
flow of blood. PET detects the locations where positrons are
that purpose although there are many technical and con-
ceptual issues unresolved4.
Other attempts to bridge the divide between the two
paradigms involve computational models which aim to
explain how higher cognitive functions could possibly
be supported by a distributed architecture. In order to
achieve this, descriptive elements from different levels
are brought together in an attempt to present unified
and coherent CMSs of cognitive processes. In the case
of models that are based on simple feed forward ANNs
and SRNs, theoretical and conceptual elements are sub-
jected to a set of neurologically inspired mathematical
tools. An important contribution to the apparent success
of these models is that the analysis and interpretation of
experimental results can be framed in the language of
the theoretical and conceptual entities concerning the
cognitive function. Building models using ANNs is not
a difficult task, particularly if the ANN is small, because
many of the technical and methodological details need
not to be dealt with5.
A Universal Framework
Artificial neural networks are trained using algorithms
that adjust the weights between units, i.e. model neu-
rons, so that the error between the ANN’s computed
output and the expected output is minimized for the
given input. This process is repeated for all possible
input-output pairs many times over. For example, to
implement the XOR-function
On = ( I1 ∧ -I2 ) ∨ ( -I1 ∧ I2 )
the network will be presented with values for I1 and I2 ,
i.e. ‘0, 0’, ‘0, 1’, ‘1, 0’ and ‘1, 1’. The weights are adjusted
using an appropriate algorithm to minimize the error be-
tween the network’s output and the output of the train-
ing set, i.e. ‘0’, ‘1’, ‘1’, and ‘0’ respectively. Once the
network is trained, it will compute the output On from
the inputs I1 and I2 according to the XOR-function. In
many discussions about ANNs in the context of cogni-
tive modeling, the inputs are labeled with terms other
than ‘0’s or ‘1’s. Because we can use these labels freely,
there is always the danger of introducing ‘wishful’ ter-
minology not only for labels, but also for methodological
terms. But, as Fodor and Pylyshyn (1988) have pointed
out,
[. . . ] the labels play no role at all in determining
the operation of a Connectionist machine; in par-
ticular, the operation of the machine is unaffected
by the syntactic and semantic relations that hold
among the expressions that are used as labels. To
emitted from decaying atoms of a radioactive tracer (typically
H2O15), while fMRI detects different levels of oxygenated and
deoxygenated hemoglobin.
4See for example Uttal (2001) for a discussion of issues
surrounding current methods in mapping cognitive functions
onto the brain.
5 Many computer programs that implement various ANNs
are freely available, and little technical expertise is required
to use them.
1185