2.

The Human
Information-Processor
2 .1 . THE MODEL HUM AN PROCESSOR
The Perceptual System
The Motor System
The Cognitive System
2 .2 . HUM AN PERFORMANCE
Perception
Motor Skill
Simple Decisions
Learning and Retrieval
Complex Information-Processing
2 .3 . CA VEATS AND C O M PLEXITIES

Our purpose in this chapter is to convey a version of the existing

psychological science base in a form suitable for analyzing human-
computer interaction. To be practical to use and easy to grasp, the
description must necessarily be an oversimplification of the complex and
untidy state of present knowledge. Many current results are robust, but
second-order phenomena are almost always known that reveal an
underlying complexity: and alternative explanations usually exist for
specific effects. An uncontroversial presentation in these circumstances
would consist largely of purely experimental results. Such an approach
would not only abandon the possibility of calculating parameters of
human performance from the analysis of a task, but would also fail in the
primary purpose of giving the reader knowledge in a form relatively easy
to assimilate.
Our tack, therefore, is to organize the discussion around a specific,
simple model. Though limited, this model allows us to give, insofar as
possible, an integrated description of psychological knowledge about
human performance as it is relevant to human-computer interaction.

23
24 2. THE HUMAN INFORMATION-PROCESSOR

2 .1 . THE MODEL HUMAN PROCESSOR

A computer engineer describing an information-processing system at

the systems level (as opposed, for instance, to the component level)
would talk in terms of memories and processors, their parameters and
interconnections.^ By suppressing detail, such a description would help
him to envision the system as a whole and to make approximate pre­
dictions of gross system behavior.
The human mind is also an information-processing system, and a
description in the same spirit can be given for it. The description is
approximate when applied to the human, intended to help us remember
facte and predict user-computer interaction rather than intended as a
statement of what is really in the head. But such a description is useful
for making approximate predictions of gross human behavior. We
therefore organize our description of the psychological science base
around a model of this sort To distinguish the simplified account of the
present model from the fuller psychological theory we would present in
other contexts, we call this model the Model Human Processor.
The Model Human Processor (see Figures 2.1 and 2.2) can be
described by (1) a set of memories and processors together with (2) a set
of principles, hereafter called the “principles of operation.” O f the two
parts, it is easiest to describe the memories and processors first, leaving
the description of the principles of operation to arise in context.
The Model Human Processor can be divided into three interacting
subsystems: (1) the perceptual system, (2) the motor system, and (3) the
cognitive system, each with its own memories and processors. The
perceptual system consists of sensors and associated buffer memories, the
most important buffer memories being a Visual Image Store and an
Auditory Image Store to hold the output of the sensory system while it is
being symbolically coded. The cognitive system receives symbolically
coded information from the sensory image stores in its Working Memory
and uses previously stored information in Long-Term Memory to make
decisions about how to respond. The motor system carries out the
response. As an approximation, the information processing of the human
will be described as if there were a separate processor for each
subsystem: a Perceptual Processor, a Cognitive Processor, and a Motor

For a survey o f computing systems in these terms see Siewiorek, Bell, and Newell
(1981).
 2.1. THE MODEL HUMAN PROCESSOR 25

Processor. For some tasks (pressing a key in response to a light) the

human must behave as a serial processor. For other tasks (typing,
reading, simultaneous translation) integrated, parallel operation of the
three subsystems is possible, in the manner of three pipelined processors:
information flows continuously from input to output with a character­
istically short time tag showing that all three processors are working
simultaneously.
The memories and processors are described by a few parameters. The
most important parameters of a memory are

¡1 , the storage capacity in items,

6, the decay time of an item, and
K, the main code type (physical, acoustic, visual, semantic).

The most important parameter of a processor is

T, the cycle time.

Whereas computer memories are usually also characterized by their

access time, there is no separate parameter for access time in this model
since it is included in the processor cycle time.
We now consider each of the subsystems in more detail.

The Perceptual System

The perceptual system carries sensations of the physical world

detected by the body’s sensory systems into internal representations of
the mind by means of integrated sensory systems. An excellent example
of the integration of a sensory system is provided by the visual system:
The retina is sensitive to light and records its intensity, wave length, and
spatial distribution. Although the eye takes in the visual scene over a
wide angle, not quite a full half-hemisphere, detail is obtained only over
a narrow region (about 2 degrees across), called the fovea. The remain­
der of the retina provides peripheral vision for orientation. The eye is in
continual movement in a sequence of saccades, each taking about 30
msec to jump to the new point of regard^ and dwelling there 60~700
msec for a total duration of

Russo (1978).
 LONG-TERM MEMORY

^LTM “
MlTM =
^LTM ~ S6fHdntic

WORKING MEMORY

AUDITORY IMAGE MwM = ^ [2.5-4.1) chunks

VISUAL IMAGE
STORE STORE 7 (5 -9 1chunks
*^WM ~ ^ [5~226| SBC
= 200 (70-1000] msec ^500 [900-3500] msec (1 chunk) = 73 (73-226) sec
/xy,s = 17 (7-17] letters /^*IS = 5 (4.4 -6.2( letters iSy^,^(3 chunks) = 7 (5 -3 4 ] sec
Kyis = Physical ^ kais = Physical KyvM Acoustic or Visual

Figure 2.1. The Model Human Processor—memories and

processors.
Sensory information flows into Working Memory through the Perceptual Processor.
Working Memory consists of activated chunks in Long-Term Memory. The basic
principle of operation of the Model Human Processor Is the Recognize-Act Cycle of
the Cognitive Processor (PO in Figure 2.2). The Motor Processor is set in motion
through activation of chunks in Working Memory.

26
 PO. Recognize-A ct C ycle o f the Cognitive Processor. On each cycle of the
Cognitive Processor, the contents of Working Memory initiate actions associatively
linked to them in Long-Term Memory; these actions in turn modify the contents of
Working Memory.

P1. Variable P erce p tu al Processor Rate P rinciple. The Perceptual Processor cycle
time Tp varies inversely with stimulus Intensity.
P2. Encoding S p e c ific ity P rinciple. Specific encoding operations performed on what
Is perceived determine what Is stored, and what is stored determines what retrieval
cues are effective In providing access to what Is stored.

P3. D iscrim in ation Principle. The difficulty of memory retrieval Is determined by the
candidates that exist In the memory, relative to the retrieval clues.

P4. Variable Cognitive P rocessor Rate Principle. The Cognitive Processor cycle
time is shorter when greater effort Is induced by increased task demands or
information loads; It also diminishes with practice.

P5. Fittses Law . The time to move the hand to a target of size S which lies a
distance D away Is given by:

T^pos = ¡ x f \ o g ^ { D / S ^ .5), (2.3)

where ¡¡^ - fOO [70~120] msec/bit.
P6. P o w er Law o f P ractice. The time T to perform a task on the nXh trial follows a
power law:

T„ T,n-
= (2.4)

where a = .4 [.2 ~ .6 ].

P7. U n ce rta in ty P rinciple. Decision time T increases with uncertainty about the
judgement or decision to be made:

T I^ H ,
where H is the information-theoretic entropy of the decision and
/^ = 150 [0-1 57 ] msec/bit. For n equally probable alternatives (called Hick’s Law),
H = iOQ^in -1-1). (2.8)
For n alternatives with different probabilities, /?•, of occurence,

/ / = S .p .lo g 2(1/p , + 1). (2.9)

P8. R a tio n ality P rinciple. A person acts so as to attain his goals through rational
action, given the structure of the task and his inputs of Information and bounded by
limitations on his knowledge and processing ability:

Goals + Task Operators + Inputs

+ Knowledge -♦- Process-limits Behavior

P9. Problem S p ac e Principle. The rational activity in which people engage to solve a
problem can be described in terms of (1) a set of states of knowledge, (2) operators
for changing one state into another, (3) constraints on applying operators, and (4)
control knowledge for deciding which operator to apply next.

Figure 2.2. The Model Human Processor— principles of

operation.

27
28 2. THE HUMAN INFORMATION-PROCESSOR

Eye-movement = 230 [70~700] msec ?

(In this expression, the number 230 msec represents a typical value and
the numbers in brackets indicate that values may range from 70 msec to
700 msec depending on conditions of measurement, task variables, or
subject variables.) Whenever the target is more than about 30 degrees
away from the fovea, head movements occur to reduce the angular
distance. These four parts—central vision, peripheral vision, eye move­
ments, and head movements—operate as an integrated system, largely
automatically, to provide a continual representation of the visual scene of
interest to the perceiver.

PERCEPTUAL MEMORIES
Very shortly after the onset of a visual stimulus, a representation of
the stimulus appears in the Visual Image Store of the Model Human
Processor. For an auditory stimulus, there is a corresponding Auditory
Image Store. These sensory memories hold information coded physically,
that is, as an unidentified, non-symbolic analogue to the external
stimulus. This code is affected by physical properties of the stimulus,
such as intensity. For our purposes we need not enter into the details of
the physical codes for the two stores but can instead just write:

K = physical,
Kai s = physical.

For example, the Visual Image Store representation of the number 2

contains features of curvature and length (or equivalent spatial frequency
patterns) as opposed to the recognized digit.
The perceptual memories are intimately related to the cognitive
Working Memory as Figure 2.1 depicts schematically. Shortly after a
physical representation of a stimulus appears in one of the perceptual
memories, a recognized, symbolic, acoustically-coded (or visually-coded)

Actual saccadic eye-movement times (travel -H fixation time) can vary quite
considerably depending on the task and the skill o f the observer. Russo (1978, Table 2,
p. 94) lists 70 msec as the minimum time and 230 msec as a typical time. The largest
time given by Busswell (1922, p. 31) for eye-movements in reading is 660 msec (for first-
grade children), which we round to 700 msec.
 2.1. THE MODEL HUMAN PROCESSOR 29

representation of at least part of the perceptual memory contents occurs

in Working Memory. If the contents of perceptual memory are complex
or numerous (for example, an array of letters) and if the stimulus is
presented only fleetingly, the perceptual memory trace fades, and
Working Memory is filled to capacity before all the items in the
perceptual memory can be transferred to representations in Working
Memory (for letters the coding goes at about 10 msec/letter). However,
the Cognitive Processor can specify which portion of the perceptual
memory is to be so encoded. This specification can only be by physical
dimensions, since this is the only information encoded: after being
shown a colored list of numbers and letters, a person can select (without
first identifying what number or letter it is) the top half of the Visual
Image Store or the green items, but not the even digits or the digits
rather than the letters.
Figure 2.3 shows the decay of the Visual Image Store and the
Auditory Image Store over time. As an index of decay time, we use the
half-life, defined as the time after which the probability of retrieval is less
than 50%. While exponential decay is not necessarily implied by the use
of the half-life. Figure 2.3 shows that it is often a good approximation to
the observed curves. The Visual Image Store has a half-life of about

8 = 200 [90~ 1000] msec

but the Auditory Image Store decays more slowly.

^ A least-squares fit to data estimated from figures appearing in Sperling (1960) and
Averbach and Coriell (1961) yields the following facts. The half-life o f the letters in
excess o f the memory span that subjects could report in the partial report condition o f
Sperling’s (1960) experiment was 621 msec (9-letter stimulus) and 215 msec (12-letter
stimulus). Averbach and Coriell’s (1961) experiment gives a half-life o f 92 msec (16-
letter stimulus). The typical value for S y j ^ has been set at 200 msec, representing the
middle o f these. The lower and upper bounds for 8 y j ^ are set at rounded-off values
reflecting the fastest subject in the condition with the shortest half-life and the slowest
subject in the condition with the longest half-life. The shortest half-life in these
experiments was 93 msec for Averbach and Coriell’s Subject GM (16-letter condition);
the longest half-life was 940 msec for Sperling’s Subject ROR (9-letter condition). It is
possible to have the average half-life be 92 msec, shorter than the half-life o f any
subject, because this average is computed by first taking the mean o f each point across
subjects, then computing the slope o f the best least-square fitting line in semilog
coordinates.
Figure 2.3. Time decay of Visual and Auditory Image Stores.
(a) Decay of the Visual Image Store. In each experiment, a matrix of letters was
made observable tachistoscopically for 50 msec. In the case of the Sperling
experiments, a tone sounded after the offset of the letters to indicate which row
should be recalled. In the case of the Averbach and Coriell experiment, a bar
appeared after the offset of the letters next to the letter to be identified. The
percentage of indicated letters that could be recalled eventually asymptotes to
/iw iv i*- The graph plots the percentage of letters reported correctly in excess of
/Xwivi* ^ function of time before the indicator.
(b) Decay of the Auditory Image Store. Nine letters were played to the observers
over stereo earphones arranged so that three sequences of letters appear to come
from each of three directions. A light lit after the offset of the letters to indicate
which sequence should be recalled. The graph plots the percentage of the
relevant 3-letter sequence In excess of reported correctly as a function of
time before the light was lit.

30
 2.1. THE MODEL HUMAN PROCESSOR 31

= 1500 [900-3500] msec

consistent with the fact that auditory information must be interpreted

over time. The capacity of the Visual Image Store is hard to fix precisely
but for rough working purposes may be taken to be about

¡iyis = 17 [7~17] letters

The capacity of the Auditory Image Store is even more difficult to fix,
but would seem to be around

M/4/5 “ ^ [4.4~6.2] letters ?

PERCEPTUAL PROCESSOR
The cycle time Tp of the Perceptual Processor is identifiable with the
so-called unit impulse response (the time response of the visual system to

The half-life o f the letters in excess o f the memory span that subjects could report
in the partial report condition o f Darwin, Turvey, and Crowder’s (1972) experiment was
1540 msec, which we have rounded to 5 ^ /^ = 1500 msec. The difference in decay
half-life as a function o f letter order in their experiment (963 msec for the third letter,
3466 msec for the first letter) has been rounded to give lower and upper bounds o f 900
and 3500. Other techniques have been used to obtain values for the “decay time” of the
Auditory Image Store. For example, use o f a masking technique gives estimates o f
around 250 msec full decay (Massaro, 1970), but these experiments have been criticized
by Klatzky (1980, p. 42) because they may only measure the time necessary to transmit
categorical information to Working Memory. On the other end, experiments that
measure the delay at which there is still some facilitation o f the identification o f a noisy
signal (Crossman, 1958; Guttman and Julesz, 1963) give very wide full-decay estimates:
from 1000 msec to 15 minutes!

^ Sperling (1963, p. 22) estimates the capacity o f the Visual Image store in terms o f
the number o f letters available at least 17 letters and possibly more. The fewest number
o f letters available for any subject immediately after stimulus presentation in the 9-letter
condition (Sperling, 1960) was 7.4 letters for Subject NJ.
Range is from the number o f letters or numbers that could be reported by
Darwin, Turvey, and Crowder’s (1972) subjects in an experiment in which they had to
give the trio o f letters coming from one o f three directions (indicated by a visual cue
shortly after the end o f the sounds). Lowest value, 4.4 letters, is for accuracy o f recalling
second letter o f triple when subjects had to name all items coming from a certain
direction (Figure 1, p. 259). Highest number, 6.2 letters, is for recall by category when
no location was required (Figure 2(B), p. 262).
32 2. THE HUMAN INFORMATION-PROCESSOR

a very brief pulse of light)* and its duration is on the order of

Tp = 100 [50- 200] msec

If a stimulus impinges upon the retina at time r = 0, at the end of time t

= Tp the image is available in the Visual Image Store and the human
claims to see it. In truth, this is an approximation, since different infor­
mation in the image becomes available at different times, much as a
photograph develops.^® For example, movement information and low
spatial frequency information are available sooner than other information.
A person can react before the image is fully developed or can wait for a
better image, according to whether speed or accuracy is the more
important.
Perceptual events occurring within a single cycle are combined into a
single percept if they are sufficiently similar. For example, two lights
occurring at different nearby locations within 60-100 msec combine to
give the impression of a single light in motion. A brief pulse of light,
lasting t msec with intensity /, has the same appearance as a longer pulse
of less-intense light, provided both pulses last less than 100 msec, giving
rise to Bloch’s Law (1885):

I»t = k, / < Tp .

Two brief pulses of light within a cycle combine their intensities in a

more complicated way, but still give a single percept.^^ Thus there is a
basic quantum of experience; and the present is not an instantaneous
dividing line between past and future, but has itself duration.
Figure 2.4 shows the results of an experiment in which subjects were
presented with a rapid set of clicks, from 10 to 30 clicks per second, and
were asked to report how many they heard. The results show that they
heard the correct number when the clicks were presented at 10 clicks/sec,
but missed progressively more clicks at 15 and 30 clicks/sec. A simple

8 See Ganz (1975).

The source o f the range is the review by Harter (1967), who also discusses the
suggestion that the cycle time can be identified with the 77~125 msec alpha period in
the brain.
10
See Ericksen and Shultz (1978), Ganz (1975).
11 See Ganz (1975).
 2.1. THE MODEL HUMAN PROCESSOR 33

Figure 2.4. Fusion of clicks within 100 msec.

A burst of sound containing an unknown number of auditory clicks at the uniform
rate of 10/sec, 15/sec, or 30/sec was presented to the subject. The graph plots
the number of clicks/burst reported as a function of the number presented. After
Cheatham and White (1954, Figure 1, p. 427).

analysis in terms of the Model Human Processor shows why. When the
experimenter plays the clicks at 10 clicks/sec, there is one click for each
Tp ~ 100 msec interval and the subject hears each click. But when the
experimenter plays the clicks at 30 clicks/sec, the three clicks in each 100
msec cycle time are fused into a single percept (perhaps sounding a little
louder) and the subject hears only one click instead of three, or 10
clicks/sec. The data in Figure 2.4 show that the number of clicks/sec
perceived by the subjects does in fact stay approximately constant in the
10 clicks/sec range (the measured values of the slopes are 9~11 clicks/
sec) for the three rates of presentation.
As a second-order phenomenon, the processor time Tp is not com­
pletely constant, but varies somewhat according to conditions. In
particular, Tp is shorter for more intense stimuli, a fact derivable from a
more detailed examination of the human information-processor using
linear systems theory, but which we simply adopt as one of the principles
of operation (Figure 2.2);
34 2. THE HUMAN INFORMATION-PROCESSOR

PI. Variable Perceptual Processor Rate Principle. The Per­

ceptual Processor cycle time Tp varies inversely with stimulus
intensity.

The effect of this principle is such that Tp can take on values within the
50~200 msec range we have given. Under very extreme conditions of
intense, high-contrast stimuli or nearly invisible, low-contrast stimuli, Tp
can take on values even outside these ranges.

The Motor System

Let us now consider the motor system. Thought is finally translated

into action by activating patterns of voluntary muscles. These are
arranged in pairs of opposing “agonists” and “antagonists,” fired one
shortly after the other. For computer users, the two most important sets
of effectors are the arm-hand-finger system and the head-eye system.
Movement is not continuous, but consists of a series of discrete
micromovements, each requiring about

■
’’a/ “ [30~100] msec,
12

which we identify as the cycle time of the Motor Processor. The

feedback loop from action to perception is sufficiently long (200~500
msec) that rapid behavioral acts such as typing and speaking must be
executed in bursts of preprogrammed motor instructions.
An instructive experiment is to have someone move a pen back and
forth between two lines as quickly as possible for 5 sec (see Figure 2.5).
Two paths through the processors in Figure 2.1 are clearly visible: (1)
The Motor Processor can issue commands (“open loop”) about once
every = 70 msec; in Figure 2.5 this path leads to the 68 pen reversals
made by the subject in the 5 sec interval, or = 74 msec/reversal. (2)
The subject’s perceptual system can perceive whether the strokes are

12
The limit o f repetitive movement o f the hand, foot, or tongue is about 10
movements/sec (Fitts and Posner, 1967, p. 18). Chapanis, Gamer, and Morgan (1949, p.
284) cite tapping rates o f 8~13 taps/sec (38~62 movements/sec, assuming 2
movements/tap). Fox and Stansheld (1964) cite figures o f 130 msec/tap = 65
msec/movement Repetition o f the same key in Kinkead’s data (Figure 2.15Z>) averages
to 180 msec/keystroke = 90 msec/movement The scribbling rate in Figure 2.5 was 74
msec/movement We summarize these as 70 [30~100] msec/movement
 2.1. THE MODEL HUMAN PROCESSOR 35

5 sec
68 Corrections
H

20 Corrections
^C o nto ur of
Stroke Bottoms

Figure 2.5. Maximum motor output rate.

Marks made by subject moving pen back and forth between two lines as fast as
possible for 5 sec.

staying within the lines (the perception process requires Tp msec) and
send this information to the cognitive system, which can then advise (the
decision process requires msec) the motor system to issue a correction
(the motor process requires msec). Tire total time, therefore, to make
a correction using visual feedback (“closed loop”) should be on the order
of = 240 msec; in Figure 2.5, this path leads to the
roughly 20 corrections about the ruled guidelines as indicated by the
dotted line tracing the contours of the bottoms of the strokes, or (5
sec)/(20 movements) = 250 msec/movement.

The Cognitive System

In the simplest tasks, the cognitive system merely serves to connect

inputs from the perceptual system to the right outputs of the motor
system. But most tasks performed by a person are complex and involve
learning, retrieval of facts, or the solution of problems. As would be
expected, the memories and the processor for the cognitive system are
more complicated than those for the other systems.

COGNITIVE MEMORIES
There are two important memories in the cognitive system: a Working
Memory to hold the information under current consideration and a Long-
Term Memory to store knowledge for future use.
36 2. THE HUMAN INFORMATION-PROCESSOR

Working Memory. Working Memory holds the intermediate products

of thinking and the representations produced by the perceptual system.
Functionally, Working Memory is where all mental operations obtain
their operands and leave their outputs. It constitutes the general registers
of the Cognitive Processor. Structurally, Working Memory consists of a
subset of the elements in Long-Term Memory that have become
activated; this intimate association between Working Memory and Long-
Term Memory is represented in Figure 2.1 by the placement of Working
Memory inside Long-Term Memory. Although Working Memory infor­
mation can be coded in many ways, the use of symbolic acoustic codes is
especially common, related, no doubL to the great importance of verbal
materials to the tasks people frequently perform. The user of a
telephone, for example, is especially liable to dial numbers mistakenly
that sound like the numbers he has just looked up. Visual codes, if
required by the the task, are also possible (as are some other types of
codes). For purposes of the Model Human Processor we consider the
predominant code types to be

K = acoustic or visual.

It is important to distinguish the symbolic, nonphysical acoustic or visual

codes of Working Memory, which are unaffected by physical parameters
of the stimulus (such as intensity), from the nonsymbolic, physical codes
of the sensory image stores, which are affected by physical parameters of
the stimulus.
The activated elements of Long-Term Memory, which define Working
Memory, consist of symbols, called chunks, which may themselves be
organized into larger units. It is convenient to think of these as nested
abstract expressions; c h u n k i = (c h u n k 2 c h u n k s c h u n k 4), with, for
instance, c h u n k 4 = (c h u n k s c h u n k s ).^^ What constitutes a chunk is as
much a function of the user as of the task, for it depends on the contents
of the user’s Long-Term Memory. The sequence of nine letters below is
beyond the ability of most people to repeat back;

BCSBMICRA
13
It is also possible to think o f these as semantic networks, such as those in
Anderson (1980) and other recent publications. At the level o f our discussion, any of
these notations will suffice about equally well. See also Simon (1974) for a technical
definition o f chunk.
 2.1. THE MODEL HUMAN PROCESSOR 37

However, consider the list below, which is only slightly different:

CBSIBMRCA

Especially if spoken aloud, this sequence will be chunked into CBS IBM
RCA (by the average American college sophomore) and easily remem­
bered, being only three chunks. If the user can perform the recoding
rapidly enough, random lists of symbols can be mapped into prepared
chunks. A demonstration of this is the mapping of binary digits into
hexadecimal digits:

0100001000010011011001101000
0100 0010 0001 0011 0110 0110 1000
4213668

This last can be easily remembered. The coding must be done in both
directions, binary to hexadecimal and hexadecimal to binary, and takes
substantial practice before it can be carried out as part of a regular
memory-span test, but it can be done. Indeed, with extended effort, the
digit span can be increased enormously. A Camegie-Mellon University
student holds the current record at 81 decimal digits, presented at a
uniform rate of 1 digit per second.^'* This particular event occurred as
part of a psychological study, where it could be verified that all the gain
was due to elaborate recoding and immense practice in its use and
development, rather than any physiological endowment.
Chunks can be related to other chunks. The chunk r o b in , for
example, sounds like the chunk Ro b e r t . It is a subset of the chunk
BIRD, it has chunk w in g s , it can chunk f l y . When a chunk in Long-
Term Memory is activated, the activation spreads to related chunks and
to chunks related to those. As the activation spreads to new chunks, the
previously activated chunks become less accessible, because there is a
limited amount of activation resource. The new chunks are said to
interfere with the old ones. The effect of this interference is that the
chunk appears to fade from Working Memory with time (unless
reactivated), as the decay curves in Figure 2.6 show. The curves are
significantly affected by other variables, including the number of other
chunks the user is trying to remember, retrieval interference with similar

14
Ericsson, Chase, and Faloon (1980); Chase and Ericsson (1981).
38 2. THE HUMAN INFORMATION-PROCESSOR

10 20 30
Time (sec)

Figure 2.6. Working Memory decay rate.

Subject is given either one or three words or consonants to remember. He counts
backwards (preventing rehearsal) for a time and then recalls stimulus. Graph plots
proportion of items correctly recalled as a function of the time elapsed until recall
began.

chunks in Working Memory, and input and retrieval memory strategies

of the user. As a working value we take the half-life of 7 sec from the
curve in Figure 2.6, which together with other data gives

5 = 7 [5-226] sec .15

The decay parameter 6 has a wide range, because most of the

apparent decay comes about from the details of interference, as we have
noted above. But these details are difficult to analyze, so it is most
convenient to accept the range and talk in terms of decay. Since the

For three chunks, Peterson and Peterson’s (1959) data (Figure 2.6) give a half-life
o f about 5 sec. Murdock’s data (Murdock, 1961) in Figure 2.6 give a half-life o f about
7 sec for 3 words and also 9 sec for 3 consonants. On the other hand, Melton’s (1963)
data give a much longer half-life o f 34 sec. For one chunk, Murdock’s data in Figure
2.6 and Melton’s (1%3) give half-lives o f 73 sec and 226 sec, respectively.
 2.1. THE MODEL HUMAN PROCESSOR 39

decay rate is particularly sensitive to the number of chunks in the

recalled item, it is useful to record the decay rate of representative item
sizes:

8 chunk) = 73 [73-226] sec

SfynfO chunks) = 7 [5-34] sec

When people are asked to recall information a few seconds after

hearing it, they use both Working Memory and Long-Term Memory to
do so. Experimentally, these two systems have been teased apart showing
that there is a pure capacity o f Working Memory (example: number of
immediately preceding digits recallable from a long series when the series
unexpectedly stops),

P w A f - ^ [2.5-4.1] chunks

When this pure capacity is augmented by the use of Long-Term Memory,

the effective capacity o f Working Memory (example: longest
number that can be repeated back) extends to the familiar 7±2 chunks,

^ chunks P

Long-Term Memory. Long-Term Memory holds the user’s mass of

available knowledge. It consists of a network of related chunks, accessed
associatively from the contents of the Working Memory. Its contents
comprise not only facts, but procedures and history as well.
Apparently, there is no erasure from Long-Term Memory,

^LTM ~

However, successful retrieval of a chunk depends on whether associations

to it can be found. There are two reasons the attempt to retrieve a
chunk might fail: (1) effective retrieval associations cannot be found, or

Crowder (1976) reviews several methods. Estimates are Waugh and Norman
(1%5) method, 2.5 items; Raymond (1969) method, 2.5 items; Murdock (19606, 1967)
method, 3.2~4.1 items; Tulving and Colatia (1970) method, 3.3~ 3.6 items. See also
Glanzer and Razel (1974).
17 MUler (1956).
40 2. THE HUMAN INFORMATION-PROCESSOR

(2) similar associations to several chunks interfere with the retrieval of the
target chunk. The great importance of these links between particular
chunks in Long-Term Memory, that is, the semantic coding of infor­
mation, leads us to list it as the predominant code type,

Klt m = semantic.

To be stored in Long-Term Memory, information from the sensory

memories must ultimately be encoded into symbolic form: a pattern of
light and dark might be coded as the letter A, an extended pattern coded
as a system error message. When the information from Working
Memory becomes part of Long-Term Memory, the precise way in which
it and the coincident Working Memory contents were encoded deter­
mines what cues will be effective in retrieving the item later. Suppose a
user names a computer-imaging file l ig h t (as opposed to d a r k ). If he
later scans a directory listing of file names to identify which ones were
the ones he created and thinks of l ig h t (as opposed to h e a v y ), he will
not be able to recognize the file, because he will be using a different set
of retrieval cues. As a principle of operation,

P2. Encoding Specificity Principle}^ Specific encoding

operations performed on what is perceived determine what is
stored, and what is stored determines what retrieval cues are
effective in providing access to what is stored

Because of interference with other chunks in memory that are more

strongly activated by the associations used as retrieval cues, information,
despite being physically present, can become functionally lost. Stated as
a principle,

P3. Discrimination Principle. The difficulty o f memory

retrieval is determined by the candidates that exist in the
memory, relative to the retrieval cues.

Items cannot be added to Long-Term Memory directly (accordingly.

Figure 2.1 shows no arrow in this direction); rather, items in Working

18 Tulving and Thompson (1973).

 2.1. THE MODEL HUMAN PROCESSOR 41

Memory (possibly consisting of several chunks) have a certain probability

of being retrievable later from Long-Term Memory. The more associ­
ations the item has, the greater its probability of being retrieved. If a
user wants to remember something later, his best strategy is to attempt to
associate it with items already in Long-Term Memory, especially in novel
ways so there is unlikely to be interference with other items. O f course
this activity, by definition, activates more items in Long-Term Memory,
causing new items to appear in Working Memory, and use capacity. On
a paced task, where a user is given items to remember at a constant rate,
the percentage of the items recalled later increases as the time/item
increases (the probability the item will be stored in Long-Term Memory
and linked so it can be retrieved increases with residence time in
Working Memory), until the time allowed per item is of the same
magnitude as the decay time of Working Memory (after which, more
time available for study does not increase the time the item is in Working
Memory), around 8 sec/chunk = 7 sec/chunk.^^
Storing new chunks in Long-Term Memory thus requires a fair
amount of time and several Long-Term Memory retrievals. On the other
hand, Long-Term Memory is accessed on every 70 msec cognitive-
processing cycle. Thus the system operates as a fast-read, slow-write
system. This asymmetry puts great importance on the limited capacity of
Working Memory, since it is not possible in tasks of short duration to
transfer very much knowledge to Long-Term Memory as a working
convenience.

COGNITIVE PROCESSOR
The recognize-act cycle, analogous to the fetch-execute cycle of
standard computers, is the basic quantum of cognitive processing. On
each cycle, the contents of Working Memory initiate associatively-linked
actions in Long-Term Memory (“recognize”), which in turn modify the
contents of Working Memory (“act”), setting the stage for the next cycle.
Plans, procedures, and other forms of extended organized behavior are
built up out of an organized set of recognize-act cycles.
Like the other processors, the Cognitive Processor seems to have a
cycle time of around a tenth of a second:

Newell and Simon (1972, p. 793) reviews experiments that gives times o f 8~13
sec/chunk.
42 2. THE HUMAN INFORMATION-PROCESSOR

= 70 [25 ~ 170] msec 20

The cycle times for several types of tasks are given in Figure 2.7. The
times vary in the 25 ~ 170 msec/cycle range, depending on the specific
experimental phenomenon and experimental circumstances with which
one wishes to identify the cycle. We have chosen as a nominal value 70
msec, about at the median of those in Figure 2.7, but have included
within the upper and lower limits all the estimates from the figure. As
with the Perceptual Processor, the cycle time is not constant, but can be
shortened by practice, task pacing, greater effort, or reduced accuracy.

P4. Variable Cognitive Processor Rate Principle. The

Cognitive Processor cycle time is shorter when greater
effort is induced by increased task demands or information
loads; it also diminishes with practice.

The cognitive system is fundamentally parallel in its recognizing phase

and fundamentally serial in its action phase. Thus the cognitive system
can be aware of many things, but cannot do more than one deliberate
thing at a time. This seriality occurs on top of the parallel activities of
the perceptual and motor systems. Driving a car, reading roadside
advertisements, and talking can all be kept going by skilled intermittent
allocation of control actions to each task, along the lines of familiar
interrupt-driven time-sharing systems.

Sununary. This completes our initial description of the Model

Human Processor. To recapitulate, the Model Human Processor consists
of (1) a set of interconnected memories and processors and (2) a set of

on
On the fast end, memory scanning rates go down to 25 msec/item (Sternberg,
1975, p. 225, Figures 8 and 9, lower error bar for LETTERS). Michon (1978, p. 93)
summarizes the search for the "time quantum” as converging on 20~30 msec. On the
slow end, silent counting, which takes about 167 msec/item (Landauer, 1962), has
sometimes been taken as a minimum cognitive task. It has sometimes been argued (Hick
1952) that the subject in a choice reaction time experiment makes one choice for each bit
in the set of alternatives, in which case a typical value would be 153 m sec/bit (Figure
2.22). Welford (1973, in Kornblum) has proposed a theory o f choice reaction in which
the subject makes a series o f choices, each taking 92 msec. Blumenthal (1977) reviews
an impressively large number o f cognitive phenomena with time constraints in the tenth
o f a second range.
 2.1. THE MODEL HUMAN PROCESSOR 43

Rate at which an item can be matched

against Working Memory:
Digits 33 [27~39] m sec/item Cavanaugh (1972)
Colors 38 m sec/item Cavanaugh (1972)
Letters 40 [24~65] m sec/item Cavanaugh (1972)
W ords 47 [3 6 -5 2 ] m sec/item Cavanaugh (1972)
G eom etrical shapes 50 m sec/item Cavanaugh (1972)
Random forms 68 [4 2 -9 3 ] m sec/item Cavanaugh (1972)
Nonsense syllables 73 m sec/item Cavanaugh (1972)

Range = 2 7 - 9 3 m sec/item

Rate at which four or fewer objects

can be counted:
Dot patterns 46 m sec/item Chi AKIahr (1975)
3-D shapes 94 [4 0 -1 7 2 ] m sec/item Akin and Chase (1978)

R ange = 4 0 -1 7 2 m sec/item

Perceptual judgement:
92 m sec/inspection Welford'(1973)

Choice reaction time:

92 m sec/inspection Welford (1973)

153 m sec/b it Hyman (1953)

Silent counting rate:

167 m sec/digit Landauer (1962)

Figure 2.7. Cognitive processing rates.

Selected cycle times (msec/cycle) that might be identified with the Cognitive
Processor cycle time.

principles of operation. The memories and processors are grouped into

three main subsystems: a perceptual system, a cognitive system, and a
motor system. The most salient characteristics of the memories and
processors can be summarized by the values of a few parameters:
processor cycle time t , memory capacity ¡i, memory decay rate 8, and