Human Cognitive Architecture

Human Cognitive Architecture
Much more work has been carried out on human cognitive architecture than
on information structures. The term ‘‘cognitive architecture’’ refers to the
manner in which cognitive structures are organized. Cognitive structures
and their relations were either discovered or emphasized as individual
structures by various researchers since the early 1930s and have been
Evolution of human cognitive architecture 219conceptualized into a unified architecture since the early 1970s. While there
are many active research areas and controversies associated with that
architecture, there is also a substantial degree of consensus concerning its
basic outline. This section describes those aspects of human cognitive
architecture around which there is broad agreement, including a brief
history of our developing understanding of the topic.
1. Working Memory
Initially designated short-term memory (e.g., Miller, 1956), it is now more
commonly referred to as working memory (e.g., Baddeley & Hitch, 1974) to
reflect the change in emphasis from a holding store to the processing engine
of the cognitive system. Working memory is the seat of consciousness and,
indeed, can be equated with consciousness in that the characteristics of our
conscious lives are the characteristics of working memory. The most
commonly expressed attributes of working memory are its extremely limited
capacity, discussed by Miller (1956), and its extremely limited duration,
discussed by Peterson and Peterson (1959). In fact, both of these limitations
apply only to novel information that needs to be processed in a novel way.
Well-learned material, held in long-term memory suVers from neither of
these limitations when brought into working memory (Ericsson & Kintsch,
1995).
While initially conceptualized as a unitary concept, working memory is
now more commonly assumed to consist of multiple streams, channels, or
processors. For example, Baddeley (e.g., Baddeley, 1992; Baddeley & Hitch,
1974) divided working memory into a visuospatial sketch pad for dealing
with two-dimensional diagrams or three-dimensional information, a
phonological loop for dealing with verbal information, and a central
executive as a coordinating processor.
A major consequence of the limitations of working memory is that when
faced with new, high element interactivity material, we cannot process it
adequately. We invariably fail to understand new material if it is suYciently
complex. In order to understand such material, other structures and other
mechanisms must be used. Processing high element interactivity material
requires the use of long-term memory and learning mechanisms.
2. Long-Term Memory
Because we are not conscious of the contents of long-term memory except
when they are brought into working memory, the importance of this store
and the extent to which it dominates our cognitive activity tend to be hidden
from us. Given this hidden nature of long-term memory, it is not surprising
that modern research into long-term memory postdated research into
220 John Swellerworking memory. It took some time for researchers to realize that long-term
memory is not just used to recognize or recall information but rather is an
integral component of all cognitive activity, including activities such as high-
level problem solving. When solving a problem, it was previously assumed
that knowledge stored in long-term memory was of peripheral rather than
central importance. De Groot’s (1965) work on chess (first published in
1946) demonstrated the critical importance of long-term memory to higher
cognitive functioning. He demonstrated that memory of board configur-
ations taken from real games was critical to the performance of chess
masters. The significance of this finding became fully apparent with Chase
and Simon’s (1973) paper on the same topic.
3. Schemas
Knowledge is stored in long-term memory in schematic form, and schema
theory describes a major learning mechanism. Schemas allow elements of
information to be categorized according to the manner in which they will be
used. Thus, for example, we have a schema for the letter a that allows us to
treat each of the infinite number of printed and hand-written variants of the
letter in an identical fashion. Schemas first became important cognitive
constructs following the work of Piaget (1928) and Bartlett (1932). They
became central to modern cognitive theory, especially theories of problem
solving, in the 1980s. As well as the work of de Groot (1965) and Chase and
Simon (1973), Gick and Holyoak (1980, 1983) demonstrated the importance
of schemas during general problem solving, and Larkin, McDermott,
Simon, and Simon (1980) and Chi, Glaser, and Rees (1982) demonstrated
the critical role of schemas in expert problem solving. As a consequence of
this work, most researchers now accept that problem-solving expertise in
complex areas demands the acquisition of tens of thousands of domain-
specific schemas. These schemas allow expert problem solvers to recognize
problem states according to the appropriate moves associated with them.
Schema theory assumes that skill in any area is dependent on the acquisition
of specific schemas stored in long-term memory.
Schemas, stored in long-term memory, permit the processing of high
element interactivity material in working memory by permitting working
memory to treat the many interacting elements as a single element. In eVect,
the interacting elements are buried within the schema that, as discussed in
more detail later, can act as a central executive by appropriately coordin-
ating those interacting elements. As an example, anyone reading this chapter
has schemas for the complex squiggles that represent a word. Those
schemas, stored in long-term memory, allow us to reproduce and
manipulate the squiggles that constitute writing, in working memory,
Evolution of human cognitive architecture 221without strain. However, we are only able to do so after several years of
learning.
4. Automation
Everything that is learned can, with practice, become automated. After
practice, specific categories of information can be processed with decreasing
conscious eVort. In other words, processing can occur with
decreasing working memory load. As an example, schemas that permit us
to read letters and words must initially be processed consciously in working
memory. With practice they can be processed with decreasing conscious
eVort until eventually, reading individual letters and words becomes an
unconscious activity that does not require working memory capacity.
Schneider and ShiVrin (1977) and ShiVrin and Schneider (1977) demon-
strated the contrast between conscious and automated processing. In his
versions of the ACT architecture, Anderson places a heavy emphasis on
automation (e.g., Anderson & Lebiere, 1998). Kotovsky, Hayes, and Simon
(1985) demonstrated the enormous benefits of automated processing to
problem-solving skill. A problem isomorph that could be solved using
automated rules was solved 16 times more rapidly than an isomorph that
required the rules to be processed consciously. Thus, high element
interactivity material that has been incorporated into an automated schema
after extensive learning episodes can be manipulated easily in working
memory to solve problems and engage in other complex activities.
5. Coalescing of Isolated Cognitive Structures and Functions into a Unified
Cognitive Architecture
While these cognitive structures and functions are studied frequently in
isolation, they can be combined into a unified cognitive architecture.
Atkinson and ShiVrin (1968) elucidated relations between working or short-
term memory and long-term memory. In depicting the flow of information
between memory stores, they presented a cognitive architecture that is at the
core of most subsequent treatments. The cognitive architecture described
here incorporates the Atkinson and ShiVrin (1968) model along with the two
learning mechanisms, schema acquisition and automation.
All conscious processing of information consists of the manipulation
of schemas, which can act as interacting elements, in working memory.
That manipulation can result in learning, which consists of the creation
of new, higher order schemas and automation. Schemas are stored in long-
term memory. They can only be brought into working memory if they
are held in long-term memory. The primary, possibly sole, function of long-
term memory is to hold hierarchically organized schemas. The limitations of
222 John Swellerworking memory refer to its limited ability to process separate schemas
that have not been incorporated into a higher level schema. Only a very small
number of schemas can be processed and they can only be held in working
memory for a few seconds. Some schemas can consist of a huge number of
interacting elements. These interacting elements are lower level schemas.
When brought into working memory, a schema, no matter what its size, is
treated as a single element. Thus, schemas have a dual function of organizing
information in long-term memory and reducing working memory load.
Automation has a similar function of reducing working memory load. On
this analysis, the two learning mechanisms of schema acquisition and
automation both have a primary function of reducing working memory load
and so allowing a limited working memory to process large amounts of
information, providing that information has, after learning, been stored in
long-term memory in the form of automated schemas. This configuration of
cognitive structures and functions has evolved to handle the information
humans must deal with