Engineering Psychology PSY 378S

1
Engineering PsychologyPSY 378S

• University of Toronto
• Spring 2005
• L7 Absolute Judgment

2
Outline

• Absolute Judgment
• Limited Capacity
• Getting around the limit Using multiple
  dimensions
• Orthogonal vs. correlated dimensions
• Integral vs. separable dimensions
• Model for absolute judgment/classification
• Design Implications

3
Absolute Judgment

• Imagine a task where youre assigning stimuli to
  categories/levels
• Stimuli range along continuum
• Might be line lengths, tone pitches, light
  intensity, texture roughness
• Stimuli presented individually to subject in
  random order
• Vary the number of lengths, pitches
• This is called an absolute judgment task

4
Absolute Judgment

• Start with 2 pitches (hi, lo) perfect
  performance
• Go to 4, people start making errors
• Can compute HT as before
• Maximum channel capacity 2.6 bits (Millers 7
  plus or minus 2)

5
Limited Channel Capacity
2.6 bits
6
Limit Not Sensory

• Level of asymptote does not reflect sensory
  limitation
• Rather working memory (STM) limitation
• Evidence
• i) discrimination of levels of dimension is
  keen--1800 different tone pitches
• ii) limits are little affected by spacing of
  stimuli on dimension

7
Application

• Absolute judgment results useful for task in
  which worker has to sort objects into levels
  along some physical continuum
• e.g., food inspection size, color, etc., fabric
  inspection, etc.
• Asking a novice to sort into 8 levels and
  expecting them to be perfect--wont work
• Also useful for industrial design when
  considering how to code objects
• e.g., socket wrenches coded by color
•  But be careful

8
Caveat

• Appropriate assignment of conceptual continua
  with perceptual continua important
• Threat level and size of a ship should be coded
  by color and size of symbol, respectively, not
  the reverse

MIL-STD 2525
9
Uni- vs. Multi-DimensionalAbsolute Judgment

• Unidimensional judgments
• Stimuli vary along one dimension only
• Observer places stimuli into 2 or more categories
• Multidimensional judgments
• Stimuli vary along more than one dimensions
  (e.g., size and colour)
• Observer places stimuli into 2 or more categories
  spread across multiple dimensions
• Information theory can assess the consistency of
  match between the stimulus and its categorization

10
Orthogonal vs. Correlated Dimensions

• Most real-world stimuli are multidimensional
• When you consider a set of stimuli they can vary
  in different ways
• Orthogonal (Independent)
• Change along one dimension does not affect other
  dimension
• Correlated (Dependent)
• Change along one dimension accompanied by change
  along other dimension

11
Some Examples

• With orthogonal dimensions, levels of one
  dimension have no bearing on the other
• e.g., sex and eye colour
• With correlated dimensions, level of one covaries
  with (tends to constrain) the other
• e.g., height and weight, sex and weight

12
Multidimensional Judgment

• Can increase capacity by adding dimensions
• But improvement does not represent the perfect
  addition of channel capacity along two dimensions

• Egeth Pachella (1969)
• unidimensional capacity 3.4 bits (10 levels)
• multidimensional capacity 5.8 bits (57 levels)
• dimensions do not sum perfectly, some information
  is lost

0 1 2 3 4 5 6 ?
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 ?
not
0 1 2 3 4 5 6 7 8 9
13
Diminishing Returns

• Each additional independent dimensions increase
  HT but with a cost
• Diminishing returns in bits per dimension

Optimal performance
10
5
8
5.8 bits
4
7 bits
6
3
Human performance
H
4
2
T
3.4 bits
2
Bits/dimension
1
0
0
1
2
3
4
5
of combined dimensions
(increasing H
)
S
14
Correlated Dimensions

• Position colour of traffic lights are redundant
  or perfectly correlated dimensions
• (Also size sometimes)
• One can be perfectly predicted from the other
• Therefore, total HSHS for any dimension

15
Correlated Dimensions

• HT gets bigger as the number of dimensions
  increases
• Since HS remains constant, HS-HT (Hloss) gets
  less as number of dimensions increases
• More reliable information

16
Summary Orthog vs. Corr

• So adding orthogonal dimensions maximizes HT, the
  efficiency of the channel
• Can get more stuff through, but less reliably
• Adding correlated dimensions maximize security of
  channel
• Have a ceiling, but information reliable

17
Separable vs. Integral Dimensions

• With orthogonal and correlated dimensions, we are
  referring to the properties of the information in
  a stimulus
• How that stimulus is related to other stimuli
• An objective property of the world
• Not the perceived form of the stimulus
• Separable vs. integral dimensions refers to the
  way that the dimensions of a multidimensional
  stimulus interact

18
Separable vs. Integral Dimensions

• Separable each dimension perceived as
  independent of other dimension(s)
• Example color and
• fill texture of object,
• perpendicular vectors

• Integral one dimension of the stimulus affects
  perception of other dimension
• dimensions are dependent
• Examples color and
• brightness of an object,
• rectangle height width

19
Garners Method

• How can we tell if dimensions are integral or
  separable?
• Garners sorting task
• Observers sort two-dimensional (or
  multi-dimensional) stimuli into categories of a
  single dimension
• Three conditions
• Control
• Orthogonal
• Correlated (redundant)

20
Garners Sorting Task Conditions

• Control sort along each dimension while ignoring
  other dimension, which is constant
• dimensions uncorrelated
• e.g., judging height of rectangles of constant
  width or width of rectangles of constant height
• Orthogonal sort along each dimension while
  ignoring the other dimension, which varies
• dimensions uncorrelated
• e.g., judging height of rectangles while width
  varies and width of rectangles while height
  varies
• Correlatedsort along either of two dimensions
• dimensions perfectly correlated
• e.g., judging the width or height of rectangles
  of constant aspect ratio (r 1.0) or area (r
  -1.0)

21
Diagnostic Results

• Results from Garners sorting task
• With integral dimensions
• best correlated sort redundancy gain
• middle control
• worst orthogonal sort filtering cost,
  interference
• With separable dimensions
• Corr ctrl orthog

22
Examples of Integral/Separable Pairs
23
GRT A Model for Classification/Absolute Judgment

• General Recognition Theory (GRT) (Maddox
  Ashby, 1996)
• Models classification as a signal detection
  problem
• Like SDT, GRT assumes that repeated presentations
  of same stimulus leads to different amounts of
  neural activity

24
GRT A Model for Classification/Absolute Judgment
Separable Dimensions
Integral Dimensions
Correlated Sort
25
Emergent Features and Configural Dimensions

• More on integral dimensions
• Nature of pairing sometimes makes a difference
• Some pairings produce emergent features a new
  unidimensional stimulus property that results
  from combining 2 or more dimensions
• redundancy gain or orthogonal cost can depend on
  sign of correlation
• referred to as configural dimensions
• gain/cost depends on emergent feature
• e.g., rectangles height and width correlation,
  rHW

rHW -1.0, emergent feature shape
rHW 1.0
26
Design Implications

• Industrial Sorting Tasks
• Quality of sort worse if worker has to judge one
  dimension and ignore the other, which is varying
  independently
• if dimensions are integral
• e.g., if worker in textile factory has to judge
  fabric hue (color), but brightness also varies
•  But sort will be facilitated if dimensions are
  correlated
• e.g., textile worker has to judge hue and
  brightness co-varies
• so brown and grey wools may vary in their
  brightness (reflectance)--this will help 

27
Design Implications

• Shepard--orientation and circle diameter
• Pilot has to check orientation of two cockpit
  dials
• Dials were the same size
• New design requires pilot to check the dial
  orientation with different size dials
• Should there be any interference?

28
Design Implications (contd)

• Symbolic categorization sorting of information
  from displays
• Operator shown artificial symbol, which
  represents levels on two or more information
  dimensions.
• e.g., meteorologist viewing symbol on display,
  which might represent temperature, humidity, wind
  speed, altitude, etc.

29
Design Implications (contd)

• If data dimensions are correlated use integral
  dimensions if orthogonal use separable
  dimensions
• Correlated dimensions represented by integral
  displays can produce emergent features, aiding
  categorization, e.g, temp. pressure in a boiler
• Unidimensional judgment impaired by integral
  displays of uncorrelated dimensions
• Is there underlying correlation between variables
  being coded? (altitude temperature) integral
  dimensions will help

30
Dimensionality and Displays Summary

• WM limitation for unidimensional stimuli
  (capacity limitation)
• When operator classifies along multiple
  dimensions
• More information can be transmitted (than with
  unidimensional)maximizes efficiency (HT)
• Bits per dimension is less (loss increased)
• HS provides limit for correlated
  dimensionsmaximizes reliability
• When operator classifies along one dimension (but
  more than one varies)
• Integral dimensions produce a redundancy gain
  when dimensions are correlated
• And interference when dimensions are orthogonal
• Emergent features can also arise depending on
  pairing of dimensionscan be very useful display
  characteristic