Download A Neurocomputational Instructional Indicator of Working Memory

ASCS09: Proceedings of the 9th Conference of the Australasian Society for Cognitive Science
A Neurocomputational Instructional Indicator
of Working Memory Load in Cognitive Load Theory
Spyridon Revithis ([email protected])
School of Computer Science and Engineering
University of New South Wales, UNSW-Sydney, NSW 2052, Australia
William H. Wilson ([email protected])
School of Computer Science and Engineering
University of New South Wales, UNSW-Sydney, NSW 2052, Australia
Nadine Marcus ([email protected])
School of Computer Science and Engineering
University of New South Wales, UNSW-Sydney, NSW 2052, Australia
Abstract
This work expands the scope of an earlier neurocomputational
study of learning-performance aspects of human behaviour in
automated learning environments by applying metrics of
instructional efficiency within the conceptual framework of
cognitive load theory. The test bed is a simulated learning
environment, in which the organization of the learning
content is represented as a dependency graph of topics. The
human learning behaviour is neurocomputationally recorded
per completed topic in terms of externally-assessed
performance and introspective mental effort. The statistically
representative behavioural instances monitored are mapped
within the patterns of two Kohonen self-organizing maps
(SOM); PSOM is a compressed version of the human
performance space and ESOM is a compressed version of the
human introspective mental-effort space. PSOM and ESOM
are, in essence, models of learning behaviour for the specified
learning environment, and can be used to a) pinpoint
instructional design weaknesses associated with the
dependency graph of the learning content and b) indicate
working memory load levels and cognitive load
characteristics associated with instructional design. The latter
is achieved by combining mental effort and performance
behavioural patterns, an established dual-data metric referred
to as instructional efficiency, within the interpretational
framework of cognitive load theory.
Keywords: Cognitive Load; Self-Organizing Maps; Working
Memory; Instructional Design.
Introduction
Understanding how people learn and acquire knowledge has
been a focal point in education for at least the last three
decades, as there is an acute demand for the efficient
dissemination and assimilation of knowledge (Sweller,
1999). Although information technology has long been
applied to education, there is still a technological gap
between learning behaviour and instructional design. In
recent years attempts have been made to construct
intelligent adaptive systems to monitor student learning (e.g.
Ben-Naim, Marcus, & Bain, 2007), part of a spectrum of
studies based on Artificial Intelligence methodologies and
Web engineering (Wenger, 1987; Brusilovsky, 1999; Shang,
Article DOI: 10.5096/ASCS200945
Shi, & Chen, 2001), but there are standing issues in
measuring instructional efficiency.
The present work has been partially based on a relevant
earlier study on modelling human learners (Revithis, 2001),
and strengthens the applicability prospects of autonomous
Internet-based learning environments (Shi, Revithis, &
Chen, 2002). It introduces a novel approach to assessing
instructional design, based on collective (as opposed to
individual) working memory load, which is linked to
cognitive load theory (Sweller, 1994; Sweller, van
Merriënboer, & Paas, 1998).
In the first part of this paper the characteristics of learning
environments are discussed in relation to the human
cognitive system, and the importance of working memory
load is analysed. In the second part a learner behavioural
model based on a dual SOM neural network is introduced
and described in terms of its usage as a collective cognitive
load indicator. The last part is devoted to a short discussion
on some of the model’s implementation details and on
pointers for future work.
Situated Learning
There is a link between human cognitive architecture and
learning behaviour from the standpoint of instructional
design. Learning environments largely determine the
functional specifications for the acquisition of knowledge as
is discussed next.
Learning Environments and Human Cognitive
Architecture
A generic learning environment is comprised of a varying
number of learners and an instructor. The learning goal,
within this type of environment, is the acquisition of the
predetermined learning content (ultimately, knowledge) by
the learners; this is principally facilitated via a system of
interactions between the learners and the instructor.
As a convention, such an environment is considered
equivalent to a unit of study (e.g. introductory algebra) and
its learning content can be introduced in a series of sessions,
in each of which a thematically-distinct topic is presented to
298
ASCS09: Proceedings of the 9th Conference of the Australasian Society for Cognitive Science
the learners (e.g. mathematical equation preliminaries,
mathematical operations and their properties, solving a
quadratic equation, etc.) The structure of the learning
content can be depicted in a dependency graph of topics.
Each topic contains part of the learning content, and has a
number (greater or equal to zero) of prerequisite topics. The
network of prerequisites is dictated by the knowledge
required prior to engaging in a certain topic; it would be a
violation of the dependency graph to engage in a topic
without having successfully completed all prerequisite
topics, which would suggest lack of prerequisite knowledge
during acquisition of new knowledge (Figure 1). A topic is
successfully completed by a learner if the learner has
performed adequately in the topic assessment.
Topic S
Topic J
Topic B
Topic Z
Topic Q
Figure 1: Example of a dependency graph of a five-topic
learning content of a learning environment.
Each topic has a high level of content self-sufficiency, so
that a learner can focus his or her efforts on it without being
forced to attend to multiple topics simultaneously. While
topics are self-sufficient they are often not independent as
certain topics require knowledge either found in another
topic or distributed across multiple topics. This situation,
resulting in the formation of a network of contentdependencies depicted in the dependency graph, dictates a
specific ordering within the learning environment and
suggests its structure for instructional design.
Every human who enters a learning environment, has
certain prior knowledge (i.e. background knowledge).
According to schema theory (Bartlett, 1932; Minsky, 1975;
Bobrow & Winograd, 1977; Larkin, McDermott, Simon, &
Simon, 1980; Chi, Glaser, & Rees, 1982; Mayer, 1992)
knowledge is organized into lossy-compressed, retrievaloriented, categorical, long-term, memory units, called
schemas, and plays a particularly critical role in the
acquisition and application of new knowledge (Sweller, van
Merriënboer, & Paas, 1998). In general, schema knowledge
is instrumental in interpreting new knowledge when
referenced by working memory.
Whether background knowledge is sufficient or not for
the environment is usually determined via some type of
screening (evaluation test) at enrolment. However, a number
of learners with an insufficient knowledge base might
eventually enter the environment without being alerted, due
to imperfections of the screening mechanism. Such a
Article DOI: 10.5096/ASCS200945
knowledge mismatch leads to incompatibility of the
learners’ existing schemas with the knowledge content of
the environment or lack of necessary schemas.
Typically, the learners go through systematic evaluation
so that the instructor can obtain qualitative information on
their acquired knowledge. The overall learner’s feedback
provides uniquely valuable information on the effectiveness
and efficiency of the instructional design of the learning
environment.
Working memory load considerations
The set of all possible sequences for successfully
completing all the topics, based on the constraints of the
dependency graph, forms the environment space. As the
number of topics increases the cardinality of the
environment space becomes explosively large1. The learner
needs to explore that space in the most beneficial way,
having to select the best completion path (i.e. traversing the
environment, from topic to topic, in such an ordering so that
a high degree of knowledge acquisition is achieved).
Knowledge acquisition can be impaired in any
educational setting. This study assumes an empirically
supported explanation of why this could happen, which is
based on working memory load (Miyake & Shah, 1999;
Kalyuga, 2006). Learners receive information primarily via
their audio-visual sensory system into their sensory memory
and then process it while it is temporarily retained in
working memory; such processing employs schemaencoded knowledge residing permanently in long-term
memory (Sweller, van Merriënboer, & Paas, 1998). These
fundamental elements of human cognitive architecture point
to working memory as a focus of conscious mental
functioning. In such an architecture, the impact of cognitive
load on knowledge acquisition is substantial.
The capacity of working memory is affected by both the
natural constraint of no more than a few chunks2 of
information being present at any time (Miller, 1956; Cowan,
2001), and the contents of the knowledge base residing in
long-term memory in the form of schemas. Straining
working memory can be due to a combination of working
memory capacity overload, and lack of compatible or
automated schemas in long-term memory (Sweller, van
1
Theoretically, if S(x) is the cardinality of an environment space
of x topics, then lim S(x) x + = x!. In Algorithmic theory
(Cormen, Leiserson, Rivest, & Stein, 2009) the dependency
graph is a directed acyclic graph (DAG), which is potentially
sparse. Each topic in the dependency graph corresponds to a
node in the DAG, and each dependency is represented by an
edge; then, each way to complete all the topics is a topological
ordering of the DAG. If the DAG is very sparse the complexity
of the cardinality of the topological ordering set is generally
O(n!), where n is the number of topics in the learning
environment. Practically, the number of dependencies is
sufficiently large, therefore the complexity bound can drop, in
the worst case, to o(n!).
A chunk of information is a piece of information, which is being
handled by working memory as a unit (i.e. a single element),
instead of a group, or a set, of disparate pieces.
→
2
∞
299
ASCS09: Proceedings of the 9th Conference of the Australasian Society for Cognitive Science
Merriënboer, & Paas, 1998). Such a capacity overload
imposed on working memory, which can severely impair
knowledge acquisition and learning, has been specified as
high cognitive load, and plays an instrumental role in
cognitive load theory (Sweller, 1994). Schema-deficient
incoming information can cause, in effect, a communication
breakdown between working memory and long-term
memory since there are no schemas in long-term memory to
explicate the new knowledge (schema referencing failure);
similarly, incoming information with a very high degree of
complexity3 can cause working memory to overload. In both
cases, the consequence can be viewed as a form of working
memory capacity virtual shrinkage (Figure 2).
Working
Memory
Long-Term
Memory
Sensory memory
Perceivable
information
Memo
Schema-deficient/Complex information
Schema-efficient information
Working memory capacity
Schema referencing failure
Scheme referencing/retrieval
Figure 2: Knowledge acquisition as a function of working
memory load and schema-based knowledge. Although the
physical capacity of working memory remains unchanged, it is practically reduced when retaining schemadeficient or highly complex information.
The Behavioural Learner Model
Human behaviour has unlimited variability and scope, but it
is naturally bounded, albeit moderately, when associated to
a specific environment, as in the case of a learning
environment, since only certain behaviour is meaningful in
any particular environment. As will be discussed in the next
section, a form of mild behavioural-space reduction, which
takes place naturally in situated learning, is an initial
working assumption in modelling learning behaviour in this
study.
In a learning environment, the instructor decides whether
the learning content is cognitively accessible to, and
instructionally appropriate for the learners. The instructor
3
Information complexity is defined here in terms of element
interactivity; that is, the degree of semantic dependence amongst
the constituents of information (Sweller, 1994). A relational
complexity metric proposes a compatible and rigorous
neurocomputational approach on information complexity and
working memory processing capacity (Halford, Wilson, &
Phillips, 1998).
Article DOI: 10.5096/ASCS200945
has to examine the totality of the learners’ feedback and use
a decision mechanism to determine the degree of knowledge
acquisition currently offered within the environment.
Situated behaviour within a learning environment can
provide information to construct a model of the learners’
performance. Furthermore, it can provide sufficient
information to build a behavioural learner model, allowing
for the assessment of instructional design and working
memory load as a function of cognitive strain and
performance during knowledge acquisition.
The model introduced in this study is based on
performance and introspective metrics in order to
quantitatively characterize the learners’ behaviour and
estimate their cognitive load. The behavioural learner
model, which fits the characterization of a behavioural
outcome model (Sun & Ling, 1997), is implemented as a
computational dual self-organizing neural network, the
characteristics and architecture of which will be presented
next.
Architecture and Function of the Model
Computational modelling offers a powerful way of
interpreting and understanding cognitive phenomena and
human behaviour (Sun, Coward, & Zenzen, 2005;
Scarborough & Sternberg, 1998; Polk & Seifert, 2002).
Neurocomputational cognitive models are based on
neurologically inspired computational systems, known as
artificial neural networks (or connectionist networks), which
allow reliable interpretations of human behaviour within the
realm of cognitive psychology (Parks, Levine, & Long,
1998; McLeod, Plunkett, & Rolls, 1998; Thomas &
Karmiloff-Smith, 2003; Elman, 2005). One type of such
modelling network is the class of self-organizing feature
maps (Willshaw & von der Malsburg, 1976; Kohonen,
1982; Kohonen, 2001).
In particular, the Kohonen self-organizing neural network
(SOM) possesses a range of powerful statistical
characteristics, including approximation of the input space,
topological ordering, density matching, and feature selection
(Haykin, 2009). It employs a Hebbian-type (Hebb, 1949)
neural learning mechanism, the result of which resembles
topographic maps in the brain that produce comparable
output if given the same input (Blasdel & Salama, 1986;
Livingstone & Hubel, 1988; Merzenich & Kaas, 1980). In
general, Hebbian neural mechanisms exhibit “ecological
validity” (Munakata & Pfaffly, 2004) since they situate
themselves in the environment without the need for errordriven or goal-oriented interaction. The importance of the
self-organization approach to cognitive modelling is
indicated by the claim that neural self-organization is the
causal and sustainability factor of classifiable behaviour,
present from simple biological neural mechanisms through
to higher-order cognitive mechanisms (Kelso, 1995).
The behavioural learner model is based on two Kohonen
neural networks that map the learners’ behaviour within the
learning environment. The first neural network, called
PSOM, represents the learners’ cognitive performance per
300
ASCS09: Proceedings of the 9th Conference of the Australasian Society for Cognitive Science
topic, whereas the second one, ESOM, represents the
learners’ cognitive effort per topic, during their learning
process. Both behavioural maps are formed after PSOM and
ESOM have undergone training that exposes each neural
network to a sizeable data set of representative past cases of
learners who have successfully completed all topics. In the
case of PSOM, the training set consists of numerical data on
learners’ performance on topic evaluation tests, whereas, in
the case of ESOM, it consists of learners’ self-evaluation,
numerically-scaled, introspective data on the mental effort
which was required for each topic to be successfully
completed (Figure 3).
..
.
...
..
.
…
PSOM
p-Input
p-Input: Topic
performance
scores
...
…
ESOM
e-Input
e-Input: Topic
mental effort
scores
Figure 3: Structure and training of PSOM & ESOM networks: e- & p-inputs connect to all neurons in the map.
During the training session each neural network is
sequentially presented with performance (or mental effort)
vectors of numerical data (representing scores) in order for
the network to converge to a stable state according to the
machine learning and termination characteristics of the
Kohonen neural network algorithm (Haykin, 2009). The
significance and efficiency of the learners’ behavioural
mapping depends on how statistically representative of the
learner population the training set is. Following the
assumption that a mild form of behavioural-space reduction
takes place naturally in situated learning, both the
performance and mental effort behavioural spaces for any
learning environment are generally much smaller than the
vast theoretical space of human behaviour, but still
prohibitively large. However, the explosively large space of
situated learning behaviour can be compressed into a
manageable set of behavioural patterns in the form of two
topological behavioural maps; the latter maps, PSOM and
ESOM, represent statistically dominant cognitive
behavioural patterns for performance and mental effort,
respectively. Due to the SOM properties (Haykin, 2009) it is
expected that each map would be a close approximation of
the situated behavioural space it compresses provided that it
has been trained using a statistically representative subset of
that behavioural space (Figure 4).
Article DOI: 10.5096/ASCS200945
(M) Self-organizing map
(T) Representative data set
(B) Situated behavioural space
(S) Space of human behaviour
dim S > dim B > dim T > dim M
Figure 4: Decisively-large space reduction can be achieved
via training a SOM network using a statistically representative data set drawn from the behavioural space.
At the completion of the training, PSOM constitutes a
topological cognitive behavioural map of the learners’
performance in each topic of the learning environment. The
structure of PSOM is a two-dimensional lattice where each
node (neuron) retains a vector pattern of performance.
Similarly, the trained ESOM network maps the learners’
mental effort in each topic of the environment, where each
node (neuron) of its two dimensional lattice retains a vector
pattern of mental effort. The size of each lattice is primarily
a function of how many patterns of performance (or mental
effort) the model designer determines as appropriate to be
represented by the SOM network, and in the simulations it
was in the order of 100 nodes.
It is important to clarify that each node on either map
contains not the performance (or mental effort) of a single
learner in the environment, but a performance (or mental
effort) pattern; that is, it represents the centre of a cluster of
similar performances (or mental effort) for a statistically
dominant group of learners. Therefore, each node represents
a distinct family of characteristic learning behaviours drawn
from the learners’ situated behavioural space. Due to the
SOM topological property, neighbouring nodes represent
similar behaviours.
Cognitive load metrics and instructional efficiency
The formation of learner behaviour maps allows for
determinations to be made in relation to the learners’
cognitive response to the instructional design of the learning
environment. It has been shown (Revithis, 2001) that a
performance-based neural self-organizing learner model can
decisively assist in revealing behavioural patterns associated
with topic dependencies that are not depicted in the
dependency graph. For instance, learners might tend to
perform poorly if they take topics in a certain order despite
the fact that the specific ordering does not violate the graph
of known dependencies. Specifically, in the example in
Figure 5, the graph depicts a hidden dependency, according
to which Topic S is also a prerequisite of Topic J; this was
revealed by the trained PSOM behavioural patterns, in
which learners tend to perform poorly on Topic J if Topic J
is completed before Topic Z and Topic S. Lack of inclusion
of such hidden dependencies in the dependency graph can
be attributed to inefficient instructional design, provided
that all learners have sufficient background knowledge
301
ASCS09: Proceedings of the 9th Conference of the Australasian Society for Cognitive Science
before entering the learning environment. PSOM can be
decisive in determining which ordering of topics is optimal
in terms of knowledge acquisition, and which topics present
a challenge to the learners.
components of the model and detects a universally
consistent pattern cluster (i.e. grouping of patterns contained
in the lattice nodes) allowing a profile for each topic in the
environment to be constructed.
(P)
100%
Q
J
Topic S
Topic J
Memo
Topic B
75%
Topic Z
S
Topic Q
Figure 5: Example of a hidden dependency in a
dependency graph of a five-topic learning content.
Mental workload can be measured by various methods
including subjective/introspective, physiological and
performance-based methods (Wierwille & Eggemeier,
1993). Among all these methods it has been shown that the
subjective measurement of cognitive load is less intrusive,
easier to implement than the physiological measurement
approach, and just as valid (Paas, van Merrienboer, &
Adam, 1994; Marcus, Cooper, & Sweller, 1996).
Furthermore, it has been suggested that humans can easily
introspect on their mental load using a numerical scale
(Gopher & Braune, 1984), and that the scaling convention
does not affect the result of the introspection (Borg, 1978;
Hendy, Hamilton, & Landry, 1993). This evidence supports
the decision to include a subjective metric of mental effort
(ESOM) in the model.
By combining performance and introspective data in the
behavioural model the learning environment can be
evaluated in terms of cognitive load and instructional
design. The pairing of performance and mental effort scores
is aligned with suggestions from the literature that highly
recommend such a combined metric for instructional design
measurement of efficiency (Sweller, Merriënboer, & Paas,
1998; Marcus, Cooper, & Sweller, 1996).
The data (scores) recorded in the trained PSOM and
ESOM are critical in determining the learners’ cognitive
inclinations; based on such data the instructor can deduce,
among others, where the learners usually excel in terms of
knowledge acquisition, where they struggle but still achieve
well, where they fail to perform, and where they perform
adequately. Combining performance and mental effort
norms the learners’ scores can be interpreted based on
cognitive load theory and meaningful conclusions can be
reached indicating cognitive load levels.
In Figure 6 an example is presented for the case of a fivetopic learning environment. It is assumed here that the
instructor accesses the trained PSOM and ESOM
Article DOI: 10.5096/ASCS200945
B
50%
Topic
Q
J
S
Z
B
0%
Mental Effort
Low
High
Low
High
Medium-High
P Performance
E Effort (mental effort)
Tx Topic
Z
50%
Performance
High
High
Low
Low
Medium-High
100% (E)
Cognitive Load Indications
Low Extraneous/(High Extraneous + Low Intrinsic)
High Germane
Low Germane
High Extraneous, Low Germane, High Intrinsic
Medium-High Germane
Figure 6: Topic profiling in terms of cognitive load in the
context of learners’ dual scores (P, E) for a five-topic
learning environment.
Specifically, it is observed that the majority of learners
tend to achieve very well in Topics Q and J although
success in the latter topic comes after great effort. On the
other hand, they perform poorly in Topic S without much
effort but in Topic Z performance is poor in spite of a high
effort, whereas they achieve above average in Topic B
without putting in maximum effort. Assuming that these
patterns were retrieved from both cognitive maps the
instructor can proceed in the determination of a cognitive
load pattern for the environment. For example, it can be
determined for Topic J that, since both mental effort and
performance are high, the germane cognitive load is most
likely high. For Topic Q, on the other hand, it can be
determined that low mental effort combined with high
performance indicates either low extraneous cognitive load,
or high extraneous cognitive load combined with low
intrinsic cognitive load. Depending on the cognitive load
indications the instructor can make reliable assumptions
about the learners’ level of acquired knowledge and
working memory load, and pinpoint with high confidence
weaknesses in the environment in terms of instructional
design.
The interpretational diagram in Figure 7 illustrates the
organic relationship between the neurocomputational
components of the learner model and the implications for
cognitive load. Germane, extraneous and intrinsic cognitive
load levels are indicated in relation to performance and
mental effort scores in accordance with cognitive load
theory. In particular, the germane (G) and extraneous (E*)
cognitive load axes are placed in the diagram in relation to
the performance (P) and mental effort (E) axes. An intrinsic
cognitive load axis could not be depicted due to the
302
ASCS09: Proceedings of the 9th Conference of the Australasian Society for Cognitive Science
complexity of its relationship to the other components of the
diagram. Selected example points in this nested diagram are
placed on G and E* axes depicting cognitive load levels for
different combinations of performance and mental effort
scores.
(G)
(P)
Example points on G or E* depending on P & E
L.E ^ L.P  L.G
H.E ^ H.P  H.G ^ (L.E* ^ H.I) v (H.E* ^ L.I)
L.E ^ H.P  (L.E*) v (H.E* ^ L.I)
H.E ^ L.P  H.E* ^ L.G ^ H.I
M.E ^ M.P  M.G ^ M.E* ^ (M.I v L.I)
provide –revised– indicators of cognitive load for the
specified learning environment.
Implementation Remarks and Future Work
The model consists of two artificial neural networks, PSOM
and ESOM, of different semantic content (i.e. performance
and mental effort, respectively) but identical structure.
Specifically, each neuron weight vector of either neural
network has a) encoded performance (and mental effort,
respectively) in the same range of discretized real numbers
[0, 1], and b) has the same number of elements, which
corresponds to the number of topics in the learning
environment (Figure 8).
(E*)
PSOM neuron weight vector
(E)
Memo
P
E
^
v
Performance
Effort (mental effort)
AND
OR
Lowest
M. Medium
L. Low
H. High
Highest
E* Extraneous Cognitive Load
G Germane Cognitive Load
I Intrinsic Cognitive Load
Figure 7: Learning environment collective cognitive load as
a function of performance (P) and mental effort (E). Five
interpretational indicators of cognitive load are shown as
example points on two visible cross axes (G and E*).
For instance, low mental effort coupled with low
performance indicates low germane cognitive load (“cube”
example point). Similarly, high mental effort coupled with
low performance indicates high extraneous, low germane
and high intrinsic cognitive load (“doughnut” example
point). In some cases, the graphical depiction is ambiguous
due to the absence of the intrinsic cognitive load axis from
the diagram; this signifies that the cognitive load indication
is conditional on either the graphically-absent intrinsic
cognitive load level or the complexity of the relationship
between types of cognitive load. For example, low mental
effort coupled with high performance indicates either low
extraneous cognitive load (“rhomb” example point), or high
extraneous cognitive load combined with low intrinsic
cognitive load (not depicted in the diagram). Also, high
mental effort coupled with high performance indicates high
germane cognitive load (“hexagon” example point) and
either low extraneous cognitive load combined with high
intrinsic cognitive load (not depicted in the diagram) or high
extraneous cognitive load combined with low intrinsic
cognitive load (not depicted in the diagram).
It is important to note that the relationship between
performance & mental effort scores and levels of cognitive
load is independent of the validity of the learner model; this
is due to the fact that the latter is a stand-alone model of
situated learning behaviour with no reference to specific
interpretational assumptions from cognitive load theory.
Therefore, if the interpretational diagram (or framework) is
substantially revised the model can be equally employed to
Article DOI: 10.5096/ASCS200945
ESOM neuron weight vector
…
…
TS1
TS2
TSn
Topic performance scores
TS1
TS2
TSn
Topic mental effort scores
Figure 8: Neuron weight structure in PSOM and ESOM
neural networks for a learning environment of n topics.
The implementation of ESOM presents no challenges in
terms of representing mental effort using the specific neuron
structure and lattice size as has also been clearly
demonstrated in previous studies of modeling behaviour
using compatible SOM structures (Revithis, 2001; Revithis,
Wilson, & Marcus, 2006). PSOM semantic encoding, on the
other hand, poses the question of the feasibility of the neural
network to detect and represent behavioural inclinations
linked to hidden topic dependencies. However, elaborate
tests in controlled simulations were conducted using
training data sets with predetermined embedded behavioural
norms (Revithis, 2001), which demonstrated the ability of
the PSOM neural network to successfully form such
representations for the case of a five-topic learning
environment.
Future work includes the design and implementation of
empirical studies using real data in order to evaluate the
model’s performance on actual learning environments as
well as its scalability.
References
Bartlett F. (1932). Remembering: A Study in Experimental
and Social Psychology. New York: Cambridge University
Press.
Ben-Naim, D., Marcus, N. & Bain, M. (2007). Virtual
Apparatus Approach to Constructing Adaptive Tutorials.
Proceedings of 2007 International Conference on
eLearning, eBusiness, Enterprise Information Systems,
and eGovernment. CSREA Press.
303
ASCS09: Proceedings of the 9th Conference of the Australasian Society for Cognitive Science
Blasdel G. G., and Salama G. (1986). Voltage-Sensitive
Dyes Reveal a Modular Organization in Monkey Striate
Cortex. Nature, 321, 579-585.
Bobrow D. G., and Winograd T. (1977). An Overview of
KRL, a Knowledge Representation Language. Cognitive
Science, 1, 3-46.
Borg G. (1978). Subjective Aspects of Physical Work.
Ergonomics, 21, 215-220.
Brusilovsky P. (1999). Adaptive and Intelligent
Technologies for Web-based Education. In C. Rollinger,
and C. Peylo (Eds.), Kunstliche Intelligenz, Special Issue
on Intelligent systems and Teleteaching, 4, 19-25.
Chi M., Glaser R., and Rees E. (1982). Expertise in Problem
Solving. In R. Sternberg (Ed.), Advances in the
Psychology of Human Intelligence. Hillsdale, NJ:
Erlbaum.
Cormen T. H., Leiserson C. E., Rivest R. L., and Stein C.
(2009). Introduction to Algorithms, 3rd Ed. Cambridge,
MA: MIT Press.
Cowan N. (2001). The magical number 4 in short-term
memory: A reconsideration of mental storage capacity.
Behavioral and Brain Sciences, 24, 87-185.
Elman J. L. (2005). Connectionist Models of Cognitive
Development: Where Next? Trends in Cognitive Science,
9, 111-117.
Gopher D., and Braune R. (1984). On the Psychophysics of
Workload: Why Bother with Subjective Measures? The
Journal of the Human Factors and Ergonomics Society,
26, 519-532.
Halford G. S., Wilson, W. H., and Phillips S. (1998).
Processing Capacity Defined by Relational Complexity:
Implications for Comparative, Developmental, and
Cognitive Psychology. Behavioral, and Brain Sciences,
21, 803-864.
Haykin S. (2009). Neural Networks and Learning machines,
3rd Ed. Upper Saddle River, NJ: Prentice-Hall.
Hebb D. O. (1949). The Organization of Behavior. NY:
Wiley.
Hendy C. H., Hamilton K. M., and Landry L. N. (1993).
Measuring Subjective Workload: When is One Scale
Better than Many? The Journal of the Human Factors and
Ergonomics Society, 35, 579-601.
Kalyuga S. (2006). Instructing and Testing Advanced
Learners: A Cognitive Load Approach. New York: Nova
Science.
Kelso J. A. S. (1995). Dynamic Patterns: The SelfOrganization of Brain and Behavior (Complex Adaptive
Systems). MA: The MIT Press.
Kohonen T. (1982). Self-Organized Formation of
Topologically Correct Feature Maps. Biological
Cybernetics, 43, 59-69.
Kohonen T. (2001). Self-Organizing Maps, 3rd Ed. Berlin:
Springer.
Larkin J., McDermott J., Simon D., and Simon H. (1980).
Models of Competence in Solving Physics Problems.
Cog. Science, 4, 317-348.
Article DOI: 10.5096/ASCS200945
Livingstone M., and Hubel D. (1988). Segregation of Form,
Color, Movement, and Depth: Anatomy, Physiology, and
Perception. Science, 24, 740–749.
Marcus, N., Cooper, M. & Sweller, J. (1996).
Understanding Instructions. Journal of Educational
Psychology, 88, 49-63.
Mayer R. (1992). Thinking, Problem Solving, Cognition.
NY: Freeman.
McLeod P., Plunkett K., and Rolls E. T. (1998).
Introduction to Connectionist Modeling of Cognitive
Processes. Oxford University Press.
Merzenich M. M., and Kaas J.H. (1980). Principles of
Organization of Sensory-Perceptual Systems of Mammals.
NY: Academic Press.
Miller G. A. (1956). The Magical Number Seven, Plus or
Minus Two: Some Limits on our Capacity for Processing
Information. Psychological Review, 63, 81-97.
Minsky M. (1975). A Framework for Representing
Knowledge. In P. Winston (Ed.), The Psychology of
Computer Vision. McGraw-Hill.
Miyake A., and Shah P. (Eds.) (1999). Models of Working
Memory: Mechanisms of Active Maintenance and
Executive Control. Cambridge, UK: Cambridge
University Press.
Munakata Y., and Pfaffly J. (2004). Hebbian Learning and
Development. Developmental Science, 7, 141-148.
Paas F. G. W. C., van Merrienboer J. J. G., and Adam J. J.
(1994). Measurement of Cognitive Load in Instructional
Research. Perceptual and Motor Skills, 79, 419-430.
Parks R. W., Levine D. S., and Long D. L. (Eds.) (1998).
Fundamentals
of
Neural
Network
Modeling:
Neuropsychology
and
Cognitive
Neuroscience.
Cambridge, MA: MIT Press.
Polk T. A., and Seifert C. M. (Eds.) (2002). Cognitive
Modeling. Cambridge, MA: MIT Press.
Revithis S. (2001). A Case of Modeling Human Behavior in
Learning Environments. MS Thesis. MO, USA:
Computer
Engineering
and
Computer
Science
Department, University of Missouri, Columbia.
Revithis S., Wilson W. H., and Marcus N. (2006). IPSOM:
A Self-Organizing Map Spatial Model of How Humans
Complete Interlocking Puzzles. In A. Sattar, and B. H.
Kang (Eds.), AI 2006: Advances in Artificial Intelligence,
Lecture Notes in Artificial Intelligence, Vol. 4304. Berlin:
Springer.
Scarborough D., and Sternberg S. (Eds.) (1998). Invitation
to Cognitive Science, Volume 4: Methods, Models, and
Conceptual Issues. Cambridge, MA: MIT Press.
Shang Y., Shi H., and Chen S. (2001). An Intelligent
Distributed
Environment
for
Active
Learning.
Proceedings of the 10th International WWW Conference,
May 2001, Hong Kong.
Shi H., Revithis S., and Chen S. (2002). An Agent Enabling
Personalized Learning in e-Learning Environments.
Proceedings of the 1st International Joint Conference on
Autonomous Agents and Multiagent Systems (AAMAS),
Vol. 2. NY: ACM Press.
304
ASCS09: Proceedings of the 9th Conference of the Australasian Society for Cognitive Science
Sun R., and Ling C. X. (1997). Computational Cognitive
Modeling, the Source of Power, and Other Related Issues.
AI Magazine, 19, 113-120.
Sun R., Coward L. A., and Zenzen M. J. (2005). On Levels
of Cognitive Modeling. Philosophical Psychology, 18,
613-637.
Sweller J. (1994). Cognitive Load Theory, Learning
Difficulty and Instructional Design. Learning and
Instruction, 4, 295-312.
Sweller J., van Merriënboer J., and Paas F. (1998).
Cognitive Architecture and Instructional Design.
Educational Psychology Review, 10, 251-296.
Sweller, J. (1999). Instructional Design in Technical Areas.
Melbourne, ACER.
Thomas M. S. C., and Karmiloff-Smith A. (2003).
Connectionist Models of Cognitive Development,
Atypical Development and Individual Differences. To
appear in R. J. Sternberg, Lautrey, J., & Lubart, T. (Eds.),
Models of Intelligence for the Next Millennium. American
Psychological Association.
Wenger E. (1987). Artificial Intelligence and Tutoring
Systems: Computational and Cognitive Approaches to the
Communication of Knowledge, Morgan Kaufmann
Publishers.
Wierwille W. W., and Eggemeier F. L. (1993).
Recommendations for Mental Workload Measurement in
a Test and Evaluation Environment. The Journal of the
Human Factors and Ergonomics Society, 35, 263-281.
Willshaw D. J. and vor der Malsburg C. (1976). How
Patterned Neural Connections Can Be Set Up by SelfOrganization. Proceedings of the Royal Society of
London.
Citation details for this article:
Revithis, S., Wilson, W., Marcus, N. (2010). A
Neurocomputational Instructional Indicator of Working
Memory Load in Cognitive Load Theory. In W.
Christensen, E. Schier, and J. Sutton (Eds.), ASCS09:
Proceedings of the 9th Conference of the Australasian
Society for Cognitive Science (pp. 298-305). Sydney:
Macquarie Centre for Cognitive Science.
DOI: 10.5096/ASCS200945
URL:
http://www.maccs.mq.edu.au/news/conferences/2009/ASCS
2009/html/revithis.html
Article DOI: 10.5096/ASCS200945
305