Engineering Psychology PSY 378F

1
Engineering PsychologyPSY 378F

• University of Toronto
• Fall 2002
• L12 Memory and Training

2
Outline Lecture 1

• Working Memory
• Short-Term vs. Long-Term Memory
• evidence
• Baddeleys Working Memory Model
• Evidence
• WM Codes and Modalities

3
Outline Lecture 2

• Properties of Working Memory
• Duration, Capacity
• Items and Chunks
• Expertise
• RI and PI
• Running Memory Task
• Knowledge in the World
• ltBREAKgt

4
Outline Lecture 3

• Long-Term Memory and Training
• Levels of Processing
• Skill Acquisition
• Training Methods
• Transfer of TrainingMethods
• Negative Transfer
• ltGO HOME!gt

5
STM versus LTM

• We all know that we can know something, and then
  later forget it
• Often this is labeled short-term vs. long-term
  memory (STM vs. LTM)
• What is experimental evidence for STM/LTM
  distinction?
• Results from serial position experiments
• We give people a list of 20 words, one at a time
  at a certain rate, say 1 every 2 s (Dog, Shovel,
  Run, Ripple, )
• We do this again and again, maybe 10 times

6
Serial Position Curve

• Plot P(Recall) vs. serial position--position in
  the list

Primacy Effect
Recency Effect
P(Recall)
1
20
17
Serial Position
7
Effect of Interference

• Atkinson Shiffrin varied this procedure a
  number of different ways
• Participants perform arithmetic task before
  recalling list
• Recency effect reduced or eliminated (if time
  spent on arithmetic task increased)
• Early part of curve (including primacy effect)
  unaffected

Recency Effect
P(Recall)
Serial Position
8
Effect of Presentation Rate
Primacy Effect

• At two different presentation rates (1 s 2 s
  per item)--affects primacy only
• Different manipulations affect different parts of
  the curve
• Suggests different mechanisms produce different
  parts of the curve
• Justifies distinction between STS and LTS

P(Recall)
2 s
1 s
Serial Position
STS
LTS
P(Recall)
Serial Position
9
The Modal Model

• Two stores short-term (STS, STM, working memory)
• And long term (LTS or LTM)

10
Working Memory

• Baddeley
• Working memory is version of activity in STS
• Differs from STS in two ways
• 1) has a more functional emphasiswhat is STM
  for?
• 2) different components in STM

11
Working Memory

• Three components to working memory
• visuo-spatial sketchpad--maintenance and activity
  in visual-spatial domain (e.g., imagery such as
  mental rotation)
• central executive--contains resources that can be
  siphoned off to WM subsystems
• phonological store--verbal rehearsal--language
  based short-term storage and rehearsal

Visuospatial sketchpad
Central Executive
Articulatory Loop
Phonological Store
12
Working Memory

• What is evidence for 3 components?
• Interference occurs when two tasks (or task
  components) draw upon same working memory (WM)
  subsystem
• Performance degrades relative to situation where
  different WM subsystems involved

13
Working Memory
Visuospatial sketchpad
Task 1
Task 1
Good Time Sharing
INTERFERENCE!
Central Executive
Task 3
Task 2
Phonological Store
14
Working Memory

• An example is the experiment by Brooks (1968)

Task
Verbal
Spatial
Verbal
Response Method
 
Spatial
15
Brooks Experiment
 
 
16
Brooks Experiment Results

• Spatial task (Big F task) better performed with
  verbal responses
• Verbal task (quick brown fox nouns and verbs)
  better performed with spatial responses
• Why? Task and response method draw upon different
  WM components in these cases

17
Implications

• Implications
• The two subsystems of working memory are
  functionally independent susceptible to
  interference from different types of activities
• Tasks should be designed such that disruption
  does not occur

18
Implications (contd)

• Tasks that impose high loads on the visuo-spatial
  system--the sketch pad (e.g., air traffic
  control) should not be performed concurrently
  with other tasks that will also use this system
• Use the auditory-phonetic system (the
  phonological storearticulatory loop) instead
• On the other hand,
• Tasks involving heavy demands on the auditory
  phonetic system (e.g., editing text, computing
  numbers) will be more disrupted by concurrent
  voice input/output than by visual manual
  interaction (control with a mouse)

19
Kinesthetic WM

• Also evidence for kinesthetic working memory
• Separate from visuospatial (Woodin Heil, 1996)
• Used experienced rowers as participants
• Tapping own body interfered with memory for
  rowing positions, but not positions in 4 x 4
  matrix
• Implications for sports training and performance

Visuospatial sketchpad
Kinesthetic Output Component
Central Executive
Phonological Store
20
Codes and Modalities

• Is there an optimum matching between stimulus
  modality and working memory codes? Yes
• Although it is possible to employ
  auditory-spatial displays for spatial tasks,
  usually less effective than visual displays.
• Auditory modality less effective at processing
  spatial information
•  Tasks that demand verbal working memory better
  served by speech displays than by print (if not
  much verbal information to communicate).
• Auditory modality more effective at processing
  language information

21
Codes and Modalities

• Summary ltOHgt

22
Longer Communication

• With longer messages, both auditory and visual
  channels are likely to show failures of memory
• But with print, can physically prolong the
  message--makes it more effective for long
  messages
• Might want to code redundantly (use both auditory
  and visual displays) if it does not cause too
  much interference with other tasks

23
Quiz
24
Break
25
Duration of Working Memory

• BrownPeterson paradigm
• Participant presented with auditory sequence of
  letters
• Try to remember them while performing interfering
  task (counting backwards by 3s)

26
Duration of Working Memory
27
Duration of Spatial WM

• Loftus et al. (1979)
• Subjects tried to remember navigational info.
  similar to that delivered by air-traffic
  controllers.
• Moray (1986)
• Subjects were radar controllers trying to recall
  info. that had been displayed on a radar scope.
• Both researchers found same types of forgetting
  functions
• Essentially can clear out spatial WM memory in 18
  s or so
• So, transience also applicable to spatial WM
• Transience occurs both in visuospatial sketch pad
  AND in phonological store

28
Duration Affected by Number of Items

• Curves a and c represent 1 and 5 item (letter)
  sequences
• Faster decay observed with more items
• limiting case--curve d memory span-- 7 items

29
WM Explains Word Length Effect

• Component of phonological store is articulatory
  loop
• With more items to be rehearsed, there will be a
  longer delay between successive rehearsals of
  each item
• In fact, the length of items--how long the items
  take to say--decreases the capacity of working
  memory--so speed of rehearsal makes a difference

30
But What is an Item?

• We talked about an item being a letter
• In absolute judgment task, items were things like
  different line lengths
• Couldnt a word be an item?
• Lets try Brown-Peterson task again--with words
  this time
• DOG CAT BOY
• Pack in more information--three three-letter
  words contains nine letters
• Now we have 3X as much information being held

31
Chunking Revisited

• Miller addressed this question by proposing the
  concept of chunk
• chunking is grouping together items based on
  their meaning
• and so a chunk is that group
• e.g., b-i-f could be reorganized to FBI--now we
  have a chunk
• working memory capacity is 7 plus or minus 2
  chunks of information
• chunk can be letter, word, sentence

32
Chunking Revisited

• Components of a chunk need to be semantically
  tied together, typically through assn in LTM
• chunking can occur at higher levels as
  well--e.g., sentences
• London is the largest city in England (7
  words)--but maybe could associate the words
  together into a meaningful whole--into a
  superchunk
•  New York is the largest city in the United
  States
• Toronto is the largest city in Canada

33
Chunking Revisited

• Should avoid having people perform tasks
  requiring working at 7 plus or minus 2 limit
• One of the best ways to avoid capacity and decay
  limitations of working memory is to facilitate
  chunking whenever possible
• People with large working memory capacities
  typically have system for chunking numbers or
  letters so that they are meaningful (e.g., dates
  or ages), or by combining them hierarchically to
  form superchunks
• Ss with normal memory spans can get up to 80
  digits or so, using various chunking techniques
• Expertise plays a role herelong-term WM
  (Ericsson Kintsch)info in LTWM is stable, but
  accessed through temporarily active retrieval
  cues in WM

34
Chess

• Analogous to memory for chess position by masters
  and novices (Chase and Simon)
• If board position was taken from the progression
  of a reasonable game, experts recalled better
  than novices
• If board position random, no difference between
  the two groups

35
Pilot Communication

• Barnett (1989) found similar results with novice
  and expert pilots for communication exchange
• When exchanges flowed in the normal sequence,
  experts performed better, but no difference if
  exchanges in random sequence.
• Chunking--resulting in improved memory
  capacity--is a byproduct of training

36
English Experts

• Were all fluent in Englishall highly trained,
  experts
• Designers should be able to capitalize on
  language familiarity
• Coding--codes can be developed so as to
  facilitate chunking
• license plate codes--vanity plates are more
  memorable e.g., FUN2GO
• commercial phone numbers (967-1111)
• radio station codes (CHUM-FM, The Edge)

37
RI and PI

• Information can be lost from working memory
  through active interference from other
  information
• Retroactive Interferencematerial learned after
  material to be recalled (MTBR) affects recall of
  MTBR
• Proactive Interference material learned before
  MTBR affects recall of MTBR

38
RI and PI

• Not just a laboratory phenomemon
• Loftus study with air-traffic controllers
• At least 10 s delay necessary before material
  remembered in previous exchange did not disrupt
  memory for a subsequent exchange

39
Running Memory Task

• In the running memory task, a sequence of items
  (e.g., letters, numbers) is presented to the
  operator, and the operator has to identify the
  item K items ago

40
Running Memory Task

• Operator does not know how long the string is.
• Operator is not expected to remember the entire
  string.
• As each item comes in, the operator is expected
  to do something with it (categorize it, check its
  value, etc.)
• So, if operator asked to recall last few items,
  typically cant remember much more than n-2,
  where n is the most recent item. (performance
  falls off rapidly if K gt 2)

41
Yntema (1963) results

• Results
• 1) Ss performed better with a few objects, many
  attributes than with many objects, few
  attributes--integration/chunking effect
• 2) Perf. better if each attribute has its own
  scale

42
Yntema (1963) Recommendations

• From Result 1 Assign each operator to monitor
  all attributes of a few objects
• From Result 2 Dont code spatial variables with
  same units--e.g., distance from flight tower
  (feet) and altitude (feet).
• In addition to air-traffic control, results may
  be applicable to other domains where information
  isnt continuously shown (e.g., taxicab
  dispatcher)

43
Putting Memory in the World

• Knowledge is not all in the head--it is partially
  in the world, and in the constraints of the world
  (Norman, Design of Everyday Things)
•  
• Result precise behavior can result from
  imprecise knowledge for four reasons
• 1) Information is in the world
• 2) Great precision (of knowledge) not required
• 3) Natural constraints are present
• 4) Cultural constraints are present

44
1) Information is in the World
45
Information is in the World

• The info you code in memory need only be precise
  enought to sustain the quality of behavior you
  desire
• Whenever information needed to do a task is
  readily available, the need for us to learn it
  diminishes
• Examples
• penny
• hunt-and-peck typists
• I can take you there, but I cant tell you how
  to get there

46
2) Great Precision (of Knowledge) not Required

• Dont need all information in head
• Can distinguish quarter from nickel, although may
  not be able to tell you who is on each coin, or
  the words on the coins
• But if you make more precise memory necessary you
  will have a problem

47
Great Precision (of Knowledge) not Required

• US Susan B. Anthony one-dollar coin--confusable
  with quarter
• Britain one-pound coin--confusable with five
  pence piece
• France 10-franc coin confusable with half-franc
  coin
• Descriptions formed to distinguish among the old
  coins were not precise enough to distinguish
  between the new one and one of the old ones 

48
My Red Notebook

• I buy a small red notebook
• Call it my notebook
• Then get another notebook--a blue one
• Call first notebook my red notebook
• Then get a large red notebook
• Call first notebook my small red notebook
• Point is that my mental representation need only
  discriminate among the choices in front of me
• But add another choice and I have to change my
  representation

49
3) Natural Constraints Are Present

• Often an objects physical features limit how it
  can be used
• Cant use a shovel to brush teeth

50
4) Cultural Constraints Are Present

• Society has evolved many conventions that govern
  acceptable social behavior
• This lets us know what to do in unfamiliar
  circumstances
• What is appropriate behavior at a party, or in a
  restaurant
• What is the sequence of events in a restaurant?
• If we have to wait for something to happen (like
  the waitress to come and take our order, some of
  us get fidgety)

51
Tradeoff between Knowledge in the World and in
the Head

• We need both knowledge in the world and in the
  head
• But in certain situations we choose to rely more
  on one than the other
• Gaining the advantages of knowledge in the world
  means losing the advantages of knowledge in the
  head.

52
Tradeoff Examples

• Providing a visual echo of a message a pilot
  receives from air-traffic control
• or a continuous record of location in a
  hierarchical computer database
• or providing CDTIs (cockpit displays of traffic
  info.)
• All these examples are putting information in the
  world
• But causes visual clutter, might disrupt
  performance of pilot or user
• With CDTIs, may increase the visual workload--is
  the increase worth the benefits?
• Memory aids (information in the world) a mixed
  blessing
•  

53
Tradeoff between Knowledge in the World and in
the Head
From Norman (1992), Design of Everyday Things
54
Break
55
Quiz 2
56
Part 3

• Score each recalled word as Case, Rhyme, Semantic
  for Y and N answers separately
• Count up the number in each category

Case
Rhyme
Semantic
57
Levels of Processing

• More deeply you process something, the better the
  chance that you will remember it
• That is, that you will transfer the info to LTS
  from STS
• Deeper approx. equal to more meaningful
• Process view of memory

P(Recall)
Case
Rhyme
Semantic
Level of Processing
58
Levels of Processing Another Take

• Normans taxonomy of memory
• 1) Memory for arbitrary things
• 2) Memory for meaningful relationships
• 3) Memory thru explanation

P(Recall)
Relationship
Arbitrary
Explanation
Level of Processing
59
1) Memory for Arbitrary Things

• Items to be remembered are arbitrary
• No particular relationship to each other or to
  anything else
• Storage of arbitrary codes
• Requires rote learning--like learning the
  alphabet
• Rote learning creates problems
• It is difficult, can take considerable time and
  effort
• When problems arise, memorized sequence gives no
  hint as to what has gone wrong
• No suggestion of what you might do to fix the
  problem

60
2) Memory for Meaningful Relationships

• Can relate what we learn to knowledge that we
  already have
• New material can be understood, interpreted,
  integrated, with previously acquired material
•   e.g., Mr Tanakas L/R turn signals on
  handlebars
• Now much easier to interpret and remember
• Although doesnt really explain anything
• Cant be used for future prediction

61
3) Memory Thru Explanation

• Material can be derived from some explanatory
  mechanism
• Mental model scan play a role here
• Details can be derived when need, such as in
  unexpected situations
• People often make up mental models for many
  things that they do
• This is why designers should provide users with
  appropriate models
• When not supplied, people will make them up
  (e.g., impetus model)
• Power of a mental model is that it allows you to
  predict

62
Long-Term Memory and Training

• The HF practitioner is often faced with the
  problem of developing the most efficient training
  program--the greatest level of proficiency per
  dollar invested.
• Different forms of training are necessary for
  mastery of declarative knowledge vs. procedural
  knowledge

63
Declarative vs. Procedural Knowledge

• Declarative Knowledge--facts about a domain, we
  can verbalize these, or write them down (e.g.,
  knowledge in typical university course)
• better off with study and rehearsal
• Levels of Processing (both kinds) important here
• Procedural Knowledge--how to do something, often
  not easily verbalized (e.g., riding a bike,
  driving a car, using a lathe)
• Tell someone everything you know about riding a
  bike, but it wont help that much
• better off with practice and performance
• Were going to focus on training the second kind
  of knowledge (procedural)

64
Skill Acquisition

• Practice makes perfect
• Most skills continue to improve for weeks,
  months, even years!
• Can obtain errorless performance in many tasks
  quite quickly
• But two other performance measures continue to
  improve speed (RT), attention or resource demand
  (as measured by performing a concurrent task)

65
Still Improving After Millionth Cigar
66
3 Stages of Skill Acquisition

• Anderson (1982)
• i) cognitive stage
• learner often works from instructions, or an
  example
• learner rehearses instructions, e.g., driving
  std, press clutch down first
• ii) associative stage
• go from declarative rep. to procedural rep.
• Performance becomes more fluid and error free
• Verbalization goes
• iii) autonomous stage
• skill becomes more automated and rapid--less
  conscious
• person loses ability to verbally describe the
  scale
• performance overlearned

67
Production Rules

• Anderson talks about production rules (if-THEN)
• e.g., if high RPM and in first gear, THEN switch
  to second gear
• Says they are the key structure unifying course
  of skill acquisition
• Development of skill in associative stage can be
  decomposed into many component production rules
• Motor program is THEN part of production rule
  its learning in the autonomous stage is the
  fine-tuning of the production rule.
• To get automaticity, stimuli or rules must be
  consistently mapped to a response

68
Guided Training

• Practice makes permanent
• Training that allows errors to be made trial
  after trial will become detrimental, b/c errors
  become learned
• Guided training ensures that learners
  performance never strays far from what task
  requires

69
Training Wheels

• Error prevention often accomplished by guided
  training such as the training wheels idea
  developed by Carroll.
• With training wheels, users are prevented from
  straying off the beaten path--making typical
  mistakes that can result in wasted time
• Instead of allowing the error to affect the
  system, training wheels simply informs the
  user/learner of the nature of error, and allows
  the user to continue on
• Good evidence to support this approach

70
Augmented Feedback

• Error prevention can also be accomplished by
  using augmented feedback techniques
• For learning to fly, such feedback might paint
  an ideal flight path through the sky to the
  runway
• Learner tracks path to achieve proper landing
  approach-- ingrains the correct sequence of
  responding
• Does help to produce rapid learning of the skill

71
Problem for Guided Training

• Whats the problem with training wheels/augmented
  feedback?
•  Problem--it often leads to poor transfer in a
  more realistic environment.
• Sometimes making errors leads to learning
• Need happy mediumeliminate sources of error that
  change the task or waste training time
• But keeping those sources of error intrinsic to
  task

72
Adaptive Training

• Imagine learning to play piano
• Some component of the task made simpler to reduce
  the initial level of difficulty.
• Then, as training proceeds, this component
  gradually increases in difficulty until level of
  target task is reached.
•  

73
Evaluation of Adaptive Training

• Reviews mixed on this technique
• Simplification does make it easier to perform the
  consistent elements of the task
• However, the easy versions of the task may induce
  a response strategy incompatible with one
  necessary to perform the final task
• Time stress is effective in adaptive training,
  however
• increase the time stress (speed at which events
  occur) as approach the final task

74
Part-Task Training

• Elements of complex task learned separately
• Wightman Lintern (1985)--distinguished between
  two different forms
• segmentation and fractionization

75
Segmentation

• Segmentation defines situation where different
  sequential phases of the skill are practiced
  before being integrated
• e.g., train up on difficult passage alone, then
  play easy passage once, then play them together
• research shows this is useful--not wasting time
  on easy stuff--efficient

76
Fractionization

• Practice on components of a task separately
  (e.g., LH, RH on piano) that you eventually
  perform concurrently
• Merits less clear cut--prevents development of
  time-sharing skills--may be necessary to link and
  co-ordinate the two activities
•  If you are very careful in selecting components
  of tasks that can be easily broken off, and what
  must be practiced together fractionization
  training can be effective

77
Varied Priority Training

• (Gopher, Weil, Siegel)--effective
• Perform everything together, but attend to one
  component and de-emphasize any others
• Integrality of task not destroyed, and yet since
  only small attention is paid to the lower
  priority component, it does not distract from the
  main component

78
Transfer of Training

• Can learning a new skill, or a skill in a new
  environment, capitalize on what has been learned
  before?
• e.g.,
• Learning Excel then Access
• Training in a flight simulator before training in
  a plane
• Training course before on-the-job training

79
Transfer of Training

• How do we measure it?
• Control group took 10 hr. to reach criterion
• Transfer group took 8 hr. to reach criterion
• Savings ctrl time - transfer time
• 10 8 2 hours
•   Transfer savings control time
• 2/10 20

80
Transfer Effectiveness Ratio (TER)

• But wait a minute
• Control Group spent 10 hours training,
• Training Group 2 spent 12 hours training, 4 hours
  in simulator, 8 hours in real task
• The Transfer Effectiveness Ratio (TER) expresses
  this relative efficiency
• TER savings training period
•  2/4 .50

81
(No Transcript)
82
Transfer Effectiveness Ratio (TER)

• If TER gt 1, training for transfer group more
  efficient than for ctrl grp
• If TER lt 1, opp. is true
• If TER lt 0, your training program is worthless
• If 0 lt TER lt 1, this does not mean training is
  worthless, for two reasons
• 1) training program may be safer
• 2) may be less expensive

83
Training Cost Ratio (TCR)

• Training Cost Ratio (TCR) reflects the cost
  component
• TCR training cost in task environment per unit
  time / training cost in training program per unit
  time
• Cheaper the training device, the lower your
  allowable TER can be (everything else held
  constant)
• If TER ? TCR  gt 1, program is cost effective,
  otherwise not
• Even if program not cost effective, important to
  consider safety issues

84
Diminishing Returns

• There is a diminishing effectiveness of most
  training devices (as measured by the TER) with
  increased training time
• i.e., TERs decrease with time in training
• Amount of training at which TER ? TCR 1 is
  point beyond which the training program is no
  longer cost effective

85
Picking the App to Train

• Large TCR indicates potential for simulation
  training
• e.g.s importance of relative cost of training
  program vs. training in environment
• Ship navigation handling 

86
Types of Transfer

• ve transfer--training program and target task
  are highly similar
• 0 transfer--extreme differences between program
  and task
• -ve transfer--similar in some respects, different
  in others, leading to improper expectations  

87
Training System Fidelity

• Should training simulators resemble the real
  world as much as possible? NO.
• Why?
• 1) realistic simulators are expensive--added
  realism may add little to TER, but affects TCR
• e.g., plants in office situation
• 2) if similarity does not achieve complete
  identity, may lead to negative transfer
• e.g., motion does not help in flight
  simulators--isnt realistic
• 3) if high realism leads to high task complexity,
  may divert attention from critical skill to be
  learned
• hard to learn to drive a manual transmission in
  big city traffic

88
Capture Important Task Components

• Instead of total fidelity, need to understand
  which components of the target task should be
  preserved in the simulator, in the training
  situation.
• e.g., sequence of steps that user has to perform

89
Gibsons Invariants in Simulators

• Evidence for usefulness of including perceptual
  invariants
• e.g., global optical flow in flight simulator
• Optical flow in driving simulator--heading of
  vehicle relative to vanishing point
• Something that those designing virtual reality
  systems should remember
• Sense of immersion does not require extremely
  high fidelity--the task related invariants are
  what is necessary

90
Negative Transfer

• When two situations have similar stimulus
  elements but different response or strategic
  components, transfer will be negative
• Reverse position of gears for stickshift
• Displays, sound of engine these
  characteristics?will remain the same
• This is especially true if the new and old
  response are opposites

91
Negative Transfer
Stimulus Elements
Different
Same


Same
Response Elements
0
Different
92
Negative Transfer

• Negative transfer can be a concern for an
  operator who has to switch between two systems
• e.g.,
• Truck driver with two different gear arrangements
• switching from Microsoft Word to a DOS
  application (TurboC)--grabbing for the mouse
• MacIntosh, Windows consistent interface standard
• number of aircraft a pilot can fly without going
  through special training
• Two systems can be different in their display
  characteristics but can involve positive transfer
• If there is identity in response elements--e.g.,
  same ctrl movements for driving, different
  dashboard displays--get high transfer

93
End