Applying Cognitive Load Theory to Computer Science Education

In M. Petre & D. Budgen (Eds) Proc. Joint Conf. EASE & PPIG 2003
Pages 333-346
Applying Cognitive Load Theory to Computer Science Education

Dale Shaffer
Lander University
Greenwood, SC
USA

Wendy Doube
Monash University
Churchill, Victoria
Australia

Juhani Tuovinen
Charles Sturt University
Wagga Wagga, New South Wales
Australia

Abstract

Cognitive Load Theory provides a theoretical basis for understanding the learning process. It uses an
information processing model to describe how the mind acquires and stores knowledge, and to provide an
explanation for the limitations imposed by working memory.

This paper describes Cognitive Load Theory, discusses its application in a number of areas, and explores its
potential uses in understanding and improving novice programming and computer science education. A
number of research directions are suggested.


Introduction

Cognitive science deals with the mental processes behind learning, memory, and problem-solving. It
identifies, among several methods, an information-processing approach to explain how the mind works. In
the simplest form of this model, the mind is assumed to be divided into three portions, sensory, working, and
long term memory (Cooper, 1998). An adaptation of the model is shown in Figure 1, but is incomplete as it
omits all but the two most important senses for learning, sight and sound.


15th Workshop of the Psychology of Programming Interest Group, Keele UK, April 2003
www.ppig.org

---

PPIG Proposal

Page 2

Memory

Sensory memory receives stimuli from the senses, including sight, sounds, smell, taste, and touch. It is
short- lived, and if the mind is not able to identify and assign meaning to the input, the information is lost.
For example, the sense of smell can detect a particularly inviting and unidentifiable fragrance, and a few
moments later the fragrance is gone and is no longer able to be recalled in detail.

Long-term memory holds a permanent and massive body of knowledge and skills. Examples include how to
ride a bicycle, multiplication facts, how to walk, and where we live.

Working memory is the portion of our mind that allows us to think both creatively and logically, and to solve
problems. It provides our consciousness (Baddeley, 1993). Working memory is the interface between long-
term memory and sensory memory. After being filtered through sensory memory and before being stored in
long-term memory, knowledge must pass through working memory.








Sensory Memory
Working memory
auditory

sound
information
stimuli
auditory section
auditory processor


visual section
sight
visual
visual processor
stimuli
information



Long-term memory
Figure 1 A Simple Model of Memory

---

PPIG Proposal

Page 3

Working memory, however, has two limitations. Firstly, it is short-lived. Secondly, it is only capable of
processing up to approximately seven things at a time (Miller, 1956). If this capacity is exceeded, then some
or possibly all of the information is lost. Cooper (1998) gives the following example of this phenomenon.
Attempt to add the following in your head without the assistance of a calculator, pencil, or paper.


1) 46
2) 83,468,446

+ 37
93,849,937


+ 58,493,900

Most individuals are capable of performing the first addition without difficulty. However, the second
problem is almost impossible to solve mentally, even though the same process, addition, is involved. The
capacity of working memory was exceeded. If one were to employ paper and pencil for the second problem,
success would be likely since the paper formed an extension of working memory, allowing one to record
intermediate values and to not overload working memory.


Learning

Learning can be defined as the “encoding (storing) of knowledge and/or skills into long-term memory in su
ch a way that the knowledge and skills may be recalled and applied at a later time on demand” (Cooper,
1998). Introductory computer programming, like mathematics, requires declarative learning of abstract
concepts and procedural learning acquired by practice. Procedural learning retrieves declarative knowledge
into working memory, refines it, and strengthens cognitive structures. Declarative learning, in the field of
computer science, can be thought of as skills acquisition, and procedural learning can be thought of as
learning to execute those skills.

Visual knowledge appears to be encoded and processed differently from verbal knowledge (Pavio, 1990).
Learning can be increased when the same content is presented simultaneously in verbal and visual
representations in a way that facilitates associations between the two, especially if the verbal presentation is
in an aural medium and the visual medium is graphical. This suggests that memory is partitioned into an
auditory and a visual system and that working memory capacity can be expanded when both systems are
employed (Cooper, 1998). Other partitions are possible, too.

---

PPIG Proposal

Page 4

During recall, knowledge is retrieved into working memory. Retrieval plays a major role in increasing the
efficiency of the encoding and storage processes. With “rehearsal”, or repeated and regular retrieval, recall
processes gradually become “automated” and novices can progress towards expert status. Experts employ
automated retrieval processes to recall information from highly organized structures in long term memory.


Cognitive load theory

Cognitive load theory is an instructional theory developed by psychologist John Sweller to describe “the
learning process in terms of an information processing system involving long-term memory, which
effectively stores all of our knowledge and skills on a more-or-less permanent basis, and working memory,
which performs the intellectual tasks associated with consciousness” (Cooper, 1998). This theory, based on
the above view of the mind (Sweller, Van Merriënboer, & Paas, 1998), provides a model of how the mind
processes information based on two tenets:
• “Human working memory is limited; we can only keep in mind a few things at a time. This poses a
fundamental constraint on human performance and learning capacity.
• Two mechanisms to circumvent the limits of working memory are:
1) schema acquisition, which allows us to chunk information into meaningful units, and
2) automation of procedural knowledge.” (Wilson & Cole, 1996)

The first mechanism provides information about how the mind understands and processes information. For
example, one typically chunks portions of a phone number to help remember it. The phone number 235
4827 is much easier to remember than 2 3 5 4 8 2 7.

The telephone example suggests that chunking does not need an underlying meaning associated with the
elements that were chunked. One does not have a special meaning for 235, yet chunking can occur.
However, if meaning can be identified and used to define the chunks, remembering is significantly enhanced.
Consider memorizing the following chunks of information.

Vers esth atdo notte achme nnew andtou ching trut hsdon otdes erveto bere ad.

If Voltaire’s quotation is chunked differently, we would have

Verses that do not teach men new and touching truths do not deserve to be read.

---

PPIG Proposal

Page 5

The process of chunking information into meaningful units is quite similar to the way a computer
programmer would combine steps in a program into an abstraction (Shaffer, 2003). For example, the
individual programming steps of finding the sum of a series of numbers, then dividing the sum by the
number of numbers would be combined together (i.e. chunked). The abstraction has a convenient name,
average and, for those individuals able to form the abstraction, only one of the approximately seven things
that can be retained in working memory is accounted for.

Schema acquisition involves more than chunking. The information that is chunked is further processed and
placed into a schema. These hierarchical networks of information in long-term memory are used to associate
and store things in memory. For example, the schema in Figure 2 is hierarchical in that “Cars” is a higher-
order concept when compared to “Traffic lights.” However, some links between nodes are not hierarchical;
“Speed kills” and “Speed limits”, for example.

Schema formation is a very dynamic operation. Through building ever more complex schema by
assimilating portions of lower-level schemas into higher level schemas, skills are developed (Sweller, Van
Merriënboer, & Paas, 1998). Children learning to read, for example, have to build schemas for letters that
allow them to classify an infinite variety of shapes into the limited number of characters in the alphabet
(Sweller, Van Merriënboer, & Paas, 1998). This process continues throughout life as words are formed from
this collection of squiggles and meaning is interpreted from words.

The second mechanism that circumvents working memory, automation of procedural knowledge, deals with
skills acquisition. Once a particular skill is acquired, automatic processing can bypass working memory
(Sweller, Van Merriënboer, & Paas, 1998). In other words, with enough practice, an activity can be carried
out without conscious processing. For example, many individuals reach a stage in driving where their
conscious, or working, memory is free for other activities.

This automation of procedural knowledge only occurs when one has acquired a skill. Sweller, Van
Merriënboer, and Paas (1998) gives the example of solving the following algebraic equation for a.
a + b

= d
c

---

PPIG Proposal

Page 6

Individuals who have acquired the needed algebraic skill will automatically know that one needs to multiply
both sides of the equation by c, then subtract b. Some individuals, students studying elementary algebra for
example, would need to use working memory in an effort to recall the appropriate rules.

Building problem-solving skills, be it in the domain of algebraic equations or computer programming, have
at its roots automation of procedural knowledge. Once a particular problem-solving skill has been chunked
and stored in long-term memory in a way that enables prompt retrieval, the skill can be considered as being
learned.

Cognitive load is the load placed on the cognitive system by performing a specific task. It can be measured
in several indirect ways, including
• self-reporting (assigning load values on a scale),
• measurement of physiological activities (heart rate, for example),
• learner performance following treatment, and
• learner performance when a second task is performed concurrently with the primary task (Sweller,
Van Merriënboer, & Paas, 1998).





Figure 2 Portion of a schema in long-term memory (adapted from Cooper, 1998)

---

PPIG Proposal

Page 7

Cognitive load theory and learning

The extent to which curriculum materials impose a load on working memory varies widely. It “depends on
the number of elements that must be processed simultaneously in working memory, and the number of
elements that must be processed simultaneously, in turn, depends on the extent of element interactivity”
(Sweller, Van Merriënboer, & Paas, 1998).

For example, compare a student who is learning multiplication facts to one learning how to solve equations
in algebra. The student learning multiplication facts has quite a large number of facts to place into long term
memory, but most of the facts are independent of each other. However, students solving equations in algebra
have a larger load placed on working memory since they have to recognize (1) that they can divide both sides
of the equation by the same value, (2) that this process does not change the equality, and (3) that it is more
sensible to do this operation in some situations before adding a quantity to or subtracting a quantity from
both sides. The student solving algebraic equations is dealing with facts with higher element interactivity
which imposes a greater load on working memory.


Cognitive load and skills acquisition

Skills in computer programming can be classified as recurrent and non-recurrent (van Merrienboer, 1992).
Recurrent skills exhibit minimal variation in differing problem situations, for example, selection of a basic
language command. Non-recurrent skills vary from problem to problem, for example, structured
decomposition. According to van Merrienboer, instructional approaches should take these differences into
account.

Teaching in technical areas is often based on the formula of presenting a new topic, showing a few examples,
and assigning practice exercises. Some simple adjustments to the presentation-examples-practice formula
can accommodate cognitive load theory.

To learn recurrent skills, declarative knowledge of the abstract concepts on which the skills are based, as
well as procedural knowledge of the procedures or rules necessary to perform the skills, should be
simultaneously in working memory. Both the concepts and the procedures that use those concepts to solve
problems should be presented together. One approach to this is to partition the information into small

---

PPIG Proposal

Page 8

segments to prevent overload, then demonstrate application of the concepts within each segment (van
Merrienboer, 1992).

In contrast, non-recurrent skills can be taught by presenting heuristic approaches and strategies, independent
of domain reference (Robertson, 2000). For example, consider the following repetition rule.
In indeterminate looping,

use a WHILE loop when 0 iterations are possible, and

use a REPEAT UNTIL loop when at least 1 iteration is necessary.

In this situation of high element interactivity, presentation in the form of knowledge structures can aid
elaboration of existing schemas. For example, decomposition into a hierarchy of component goals or plans
is a form of knowledge structure (Figure 3).

Examples provide explicit information to facilitate initial and correct schema formation (Cooper, 1998) and
can initiate association of declarative and procedural knowledge. Following presentation, students can work
through several examples with the teacher or examine multiple problems with integrated solutions, then
immediately practise problems of the same type. By repeating this process with each type of problem,
students can build schemas, and reinforce learning by retrieving and refining those schemas. Eventually

control structures
executed more
than once
repetition
predetermined
otherwise
number of
times
minimum 0
minimum 1
for
while
repeat
until

Figure 3 Hierarchy as a form of knowledge structure

---

PPIG Proposal

Page 9

practice with a mixed set of problems “elaborates” the schema even further by strengthening associations. A
variation on this approach is to provide partially-completed examples for students to finish. Another is to
gradually reduce the amount of information and support for each task , so that as learning increases with
practice, less information is required.

One approach to practic ing recurrent skills is to decompose them into successively smaller components in the
form of domain-specific, algorithmic steps. Low element interactivity simplifies initial schema construction.
Each component is practiced individually, and then parts of the whole are practiced until the entire skill can
be integrated. For example arithmetic expressions are learned with presentation of entire expressions, then
separate presentation and examples of firstly constants, then variables followed by operations. Next
arithmetic expressions with constants and operations are practised. Finally, the three components of
constants, variables and operations in complete arithmetic expressions are practised together.

It is generally agreed by computer science educators that introductory computer science courses tend to have
high attrition rates, and that the lack of problem-solving skills is a contributing factor. This phenomena is
discussed from various viewpoints in Beaubouef , Lucas, and Howatt (2001), Shultz (2000), Roberts,
Kassianidou, and Irani (2002), Thweatt (1994), and in Wilson and Shrock (2001).

Problem-solving is usually approached using means-ends analysis. With experts who can rely on extensive
schema acquisition and formation in long-term memory, this heuristic is applied by reducing differences
between the current problem state and the goal state. This strategy seems especially effective in computer
programming.

However, novices often apply it by working backwards from the goal to the initial state, then forward to the
goal, resulting in a high cognitive load (Sweller, 1988). Cognitive load could be reduced if novices learning
to apply means-ends analysis are given problems that have no goal (Owen & Sweller, 1985). Rather than
focusing on the desired result, one is forced to focus on the initial state. Since much of computer
programming is goal-oriented, this learning approach would appear to be difficult to implement, but might
have significant value in an intermediate learning stage.

Another approach to reducing cognitive load is to increase working memory capacity by utilizing verbal and
visual channels. However, redundant information can increase cognitive load by increasing the number of
associations that have to be made between the different sources of information. (Chandler & Sweller, 1991).

---

PPIG Proposal

Page 10

One effective approach is to integrate text into an associated graphic. By placing the text directly within a
diagram, the student is not forced to split attention between multiple sources. Modern software, in the form
of algorithm animation and multimedia presentation packages where text and graphics can be added a piece
at a time, gives much promise for this approach. In addition, algorithm animation and multimedia give
students’ power over the speed in which the material is presented, and therefore redundant learning elements
can be reduced.

Bearing in mind that verbal material can be in the form of text and can place demands on visual resources if
combined with graphics, the final approach to improve learning is by providing multiple sensory input.
Partitioning information to come from visual and aural sources can further extend working memory (Pavio,
1990). Again multimedia and algorithm animation show promise in developing this approach but also show
that working memory can be overloaded by multiple media, especially if it is superfluous to the learning
content or asynchronous.

In summary, some of the possible strategies that could reduce cognitive load in introductory computer
programming courses are:

• partitioning of information into small segments, them simultaneous presentation of concepts and
procedures in demonstrations
• presentation of heuristic approaches and strategies, independent of domain reference
• organization of presentation of non-recurrent skills into knowledge structures
• presentation – repeated example and practice – broad practice
• broad practice using partially-completed examples
• gradual withdrawal of supporting information from practice tasks
• decomposition of practice tasks into small steps then gradual integration into a complete skill
• goal-free problem solving
• careful integration of text and graphics
• visual and aural sources for learning

---