Pattern Recognition

1
Chapter 2

• Pattern Recognition

2
Pattern Recognition
The process of identifying sensory patterns
3
Some examples
4
What do you see in this photograph?
5
Its a Dalmatian!
6
Whats this?
7
Is this A Frog?
8
Or a Horse?
9
(No Transcript)
10
(No Transcript)
11
Can you see the baby?
12
Is this woman smiling?
13
Is this woman smiling?
14
(No Transcript)
15
No!
16
Is this woman smiling?
17
Is this woman smiling?
18
(No Transcript)
19
Yes!
20
What are some real-live tasks or situations where
pattern recognition is important?
21
One thing is clear, pattern recognition is
heavily influenced by context
22
One thing is clear, pattern recognition is
heavily influenced by context
A couple of examples
23
Whats this?
24
(No Transcript)
25
(No Transcript)
26
What letter is this?
27
T
E
C
T
28
A classic study of the effects of context on
pattern recognition by Reicher
29
Basic Experimental Setup

• Participant is presented a letter by itself or in
  the context of other letters
• After the letter is presented, the participant
  identifies which letter had been presented by
  choosing from two alternatives.

30
Example
31
D
32

D or K ?
33
Conditions of the Experiment
34
Results
The letter is best recognized when placed in the
context of other letters that together spell a
word
35
What does result imply?
36
What does result imply?
It seems to imply that the word was recognized
before the individual letters of which is was
comprised.
37
Does this agree with common sense notions of
reading?
38
How can we account for these findings?
39
How can we account for these findings?
One way is to employ the notions of bottom-up and
top-down processing
40
Bottom-up vs. Top-down Processing

• Bottom-up processing Perceptual experience is
  built up from incoming sensory information (also
  referred to as data driven processing)
• Top-down processing Perceptual experience is
  influenced by ones knowledge and expectations
  about the world. (also referred to as
  conceptually driven processing)

41
Cognitions or Knowledge
Bottom-up (data driven) processing
Perceptions
Sensations
42
Cognitions or Knowledge
Bottom-up (data driven) processing
Perceptions
Top-down (conceptually driven processing
Sensations
43
An Example
44
Other examples?
45
Top-down processing, errors in perception,
hallucinations, the Rorshach, and other things.
46
Theories of Pattern Recognition

• Template Theory
• Feature Theory
• Structural Theory

47
Template Theory

• The stimulus pattern is compared to templates
  stored in the brain
• The stimulus pattern is recognized when it is
  matched against a template

48
To Illustrate
49
A complication How many templates would you
need to recognize all possible As?
A a A A A
A a A A A a
A
50
And how would you recognize a sloppy A?
51
And how would you recognize a sloppy A?
How could I possibly read your handwriting?
52
Could context help?
53
T
E
C
T
54
Could the template theory account for Reichers
findings?
55
Commonly cited problems with template theory

• It would seem to require too many templates to be
  able to deal with stimulus variability
• It is unclear how it would account context
  effects (like those of illustrated in Reichers
  work)

56
Feature Theory

• Stimulus patterns are recognized on the basis of
  features
• Features are the elementary components of a
  stimulus pattern

57
An Illustration
A
Is composed of
58
An Illustration
A
Is composed of
The notion is that A would be recognized on the
basis of these features being present
59
A phenomenon consistent with the notion of
feature detection
On the following slide, see how quickly you can
locate the letter O.
60
T
T
T
T
T
T
T
T
T
T
T
O
T
T
T
T
T
61
See how quickly you can find the O on the next
slide
62
Q
Q
Q
Q
Q
Q
Q
Q
Q
O
Q
Q
Q
Q
63
One of the most fruitful ways of looking at
feature detection is from the perspective of
neural networks
64
(No Transcript)
65
Incorporating Top-Down Processing
To incorporate the notion of top-down processing
we will have to add another level of analysis
66
To incorporate the notion of top-down processing
we will have to add another level of analysis
Levels of analysis feature level letter
level word level
67
(No Transcript)
68
We will also have to add the following
assumption Processing takes place at all levels
simultaneously (although logically beginning at
the lowest level first)
69
(No Transcript)
70
An analogy -- voting in the US
News Media
Voters
71
You can also incorporate inhibitory processes
72
(No Transcript)
73
Whats the advantage of feature detection?

• A finite number of feature detectors potentially
  could recognize an infinite number of patterns
• This approach is better able to account for
  stimulus variability
• This approach more easily incorporates the notion
  of top-down processing

74
Structural Theory
75
A problem with feature theory is that it doesnt
take into account the spatial relationship among
features
76
Structural theory attempts to overcome this
problem by suggesting that objects are recognized
on the basis of larger elements referred to as
geons.
77
(No Transcript)
78
Biederman suggests that all three dimensional
objects could be recognized on the basis of just
36 geons.
79
Questions for Structural Theory

• Wouldnt the theory have to address the spatial
  arrangement of geons?
• How could we distinguish between to similar
  objects that consist of the same geons in the
  same spatial arrangements (e.g., a horse and a
  cow)?
• How could we recognize something like a puddle
  which does not have recurring geons?

80
Sensory Memory(Sensory Store)
81
Sensory Memory(Sensory Store)
Sensory memory preserves sensory patterns for a
brief period time so that they can be recognized
82
Sensory Memory(Sensory Store)
Sensory memory is particularly important in
situations when the stimulus pattern is only
present for an instant.
83
Two Demonstrations of Sensory Memory

• Close your eyes, then open and shut your eyes
  quickly. Do you see a lingering image?
• Look straight ahead. Wave your index finger in
  front of your eyes. Do you see a trailing
  afterimage?

84
What are some examples when sensory memory might
be helpful to us?
85
Iconic vs. Echoic Memory

• Iconic memory Visual sensory memory
• Echoic memory Auditory sensory memory

86
Are there other forms of Sensory Memory?
87
Research on Iconic Memory by George Sperling
88
Basic Experimental Paradigm
Present matrix of letters for brief period of
time. Have participant report as many of the
letters as they can
89
Basic Experimental Paradigm
Example
90
Basic Experimental Paradigm

91
Basic Experimental Paradigm
X T H B L D Z P S N R J
92
Basic Experimental Paradigm
Recall
93
Ideally, the letter matrices in these experiments
should . . .

• Fall entirely on the fovea of the eye so that no
  eye movements are necessary
• Be presented for a sufficiently short duration
  that the person cant shift their attention
• Be presented at high contrast and followed by
  darkness

94
Ideally, the letter matrices in these experiments
should . . .

• Fall entirely on the fovea of the eye so that no
  eye movements are necessary
• Be presented for a sufficiently short duration
  that the person cant shift their attention
• Be presented at high contrast and followed by
  darkness

This typically requires the use of a
tachistoscope
95
Whole vs. Partial Report Procedures

• Whole report Participant is required to report
  the whole matrix (or as much as he or she can)
• Partial report Participant is required to
  report only a portion of the matrix

96
Whole vs. Partial Report Procedures

• Whole report Participant is required to report
  the whole matrix (or as much as he or she can)
• Partial report Participant is required to
  report only a portion of the matrix

A cue indicates which part of the matrix should
be recalled
97
Whole vs. Partial Report Procedures

• Whole report Participant is required to report
  the whole matrix (or as much as he or she can)
• Partial report Participant is required to
  report only a portion of the matrix

A cue indicates which part of the matrix should
be recalled
The cue may be presented at the same time the
letter matrix is presented, or delayed for a
short period of time
98
Illustration of Partial Report Procedure
99
Illustration of Partial Report Procedure

100
Illustration of Partial Report Procedure
L T Z M V R G H P B W Q
101
Illustration of Partial Report Procedure
Recall
102
Illustration of Partial Report Procedure
L T Z M V R G H P B W Q
In this case the letter matrix and cue are
presented simultaneously
103
Illustration of Partial Report Procedure
L T Z M V R G H P B W Q
In this case the letter matrix and cue are
presented simultaneously
On some trials the cue follows the presentation
of the matrix.
104
Illustration of Partial Report Procedure

105
Illustration of Partial Report Procedure
L T Z M V R G H P B W Q
106
Illustration of Partial Report Procedure
107
Illustration of Partial Report Procedure
Recall
108
Typical Findings with Whole Report Procedure
109
Typical Findings with Whole Report Procedure
Typically participants are only able to recall
about 1/3 of the letters in the matrix (3 to 4
letters)
110
Possible Explanations of the Whole Report Findings

• Only 1/3 of the letters registered by the visual
  system.
• All letters registered, but the participant only
  had sufficient time to process about 1/3 of the
  letters.

111
Findings with the partial report procedure help
to distinguish between alternatives

• If only 1/3 of the letters register, then the
  participant should only be able to recall 1/3 of
  the cued letters
• If all letters register, then the participant
  should recall all (or most?) of the cued letters

112
Findings with the partial report procedure help
to distinguish between alternatives

• If only 1/3 of the letters register, then the
  participant should only be able to recall 1/3 of
  the cued letters
• If all letters register, then the participant
  should recall all (or most?) of the cued letters

In other words . . .
113
If the percentage of letters recalled with
partial report exceeds that of whole report, we
can say that more letters were available than the
participant was able to recall under the
conditions of whole report
114
That implies . . .
115
That implies . . .
That when the cue is delayed with the partial
report procedure . . .
116
That implies . . .
That when the cue is delayed with the partial
report procedure . . .
And the level of performance exceeds that of
whole report . . .
117
That implies . . .
That when the cue is delayed with the partial
report procedure . . .
And the level of performance exceeds that of
whole report . . .
The information must be held in a memory system
-- Iconic Memory!
118
So lets look at what Sperling found
119
Sperlings Results
120
How to measure the duration of the icon

• Delay the cue more and more
• Determine the delay at which partial report
  performance no longer exceeds whole report
  performance
• At that delay there is no longer evidence of the
  icon
• That delay provides an estimate of how long the
  icon lasts, and hence, the duration of iconic
  memory

121
(No Transcript)
122
So, iconic memory exists (at least for vision)
and we know that its duration is about .5
sec. What other characteristics to we believe it
has?
123
Characteristics of Iconic Memory

• Duration -- about .5 sec
• Capacity -- large (more items than the individual
  can report)
• Code -- visual (or sensory) code

124
Why do we think information is coded in visual or
sensory form?
125
To understand imagine the experiment was run with
a slight modification
126

127
L 4 Z M 6 R G 3 P 5 2 Q
128
Cues High tone -- recall letters Low tone --
recall numbers
129
Results Cue will not be effective. Partial
report performance will not exceed whole report
performance
130
Why?
131
Why?
It has to do with the nature of the cue.
132
Types of cues

• Categorical cues -- require that information be
  identified and categorized
• Pre-categorical cues -- do not require
  information to be identified and categorized.

133
Types of cues

• Categorical cues -- require that information be
  identified and categorized
• Pre-categorical cues -- do not require
  information to be identified and categorized.

They are based on physical information contained
in the physical image
134
The fact that partial report performance exceeds
whole report performance only when a
pre-categorical cue is used, suggests that
information in iconic memory is in a
pre-categorical state.