Pattern Recognition
Why is pattern recognition important?
Humans ability to recognize patterns is what separates us most
from machines
Models of pattern recognition
Templates
Features
Structural analysis
The Mystery of Pattern Recognition
Templates
Templates
Match observed object to stored images
- A is recognized by matching it to stored photographlike image of previously seen A
Problems
Too many templates needed
One template for A is probably not enough
Need a template for A with a specific size, orientation,
color, etc.
Can this problem be solved with pre-processors?
Ignores intuition that objects are composed of smaller
parts
Feature Analysis
Recognize an object by breaking it down
into features
A is recognized by combining evidence for \ + - + /
Evidence for feature analysis
Neural feature detectors have been found
Simple and conjunctive feature search tasks
Simple features are detected in parallel, but combining
features requires attention to be moved across an image in a
serial manner
Asymmetries in feature search
Illusory conjunctions
Simple feature search
Look for an O
T
T
T O
T
T
T
T
T T
T
T T
T T
T
T
T
T TT T T T
T
T
T
T T T T
Simple feature search
Look for something red
T
O
T O T
T
T
O
T
O
T
O
TT O
O T T
O T
T O
O O
T O
O
T T
Conjunctive feature search
Look for a red O
O T
T O
T
T T O
T
T
T
O
T O T O
O
T O
T
T
T
O
T
O
T O T
O O
O
T
OT O T
Conjunctive feature search
Look for a red O
T O T
T
T
T O T
O
T TOT
O O
Response Time
Conjunctive Search
Pop-out
Simple feature search
Number of Stimuli in Display
Feature Integration Theory
Simple features can be detected anywhere in a
display, in parallel
Pop-out: for a simple feature search task, response time doesnt
depend on how many objects are in the display
Conjunctions of features require attention to bind
(glue) the features together
Attention must be deployed serially, to one object at a time
Attention as the force that glues otherwise free-floating features
together
For a conjunctive search task, response increases linearly with
number of objects
Disjunctive and conjunctive feature search tasks
Negative = Absent
Positive = Present
Display size = # of distractors
Conjunctive task: Target is any
object that is green AND S
Disjunctive task: Target is any
object that is green OR S
Why is disjunctive task a better
control for conjunctive task than
a simple feature search like
target is anything green?
Interpreting Slopes and Intercepts
The relation between display size
and reaction time is well described
by a line.
Slope = number of milliseconds
required per item in display
Intercept = amount of time
required if there were NO items.
Flat line = Pop out = display
size does not influence respone
time = simultaneous detection of a
simple feature everywhere within
a display
Why is the conjunctive Absent
slope twice the conjunctive
Present slope?
Why is there pop-out for
disjunctive present judgments
but not absent judgments?
Replication of Treisman & Gelade (1980)
Variables
Task: Conjunctive or Disjunctive
Target type: Present or Absent
Display size: 1 X 1, 2 X 2, 4 X 4, 6 X 6
Dependent variable: Response Time and % Correct
Predictions
Conjunctive response time > Disjunctive response time
Interaction between task and display size
Display size matters more for conjunctive than disjunctive tasks
2-to-1 slope ratio between conjunctive present and absent tasks
Interaction between task, display size, and target type?
Experimental Details
Tasks
Conjunctive
Target: Green T
Distractors: Red T, Green X
Disjunctive
Targets: Red or Green S, Blue T or Blue X (anything S OR Blue)
Distractors: Red T, Green X
Display sizes: 1 X 1, 2 X 2, 4 X 4, 6 X 6
Number of trials
400 total, so 400/(4 array sizes)/(2 tasks) =50 trials per block
50 trials = 25 present and 25 absent trials
Add jitter: 5 pixels (1 pixel = .034 cm)
Appearance: Helvetica 24
Order of 8 conditions: randomized for each participant
The Cleveland Font
picture
Key
<
>
Feature search asymmetries
It is easier to find X among Ys than Y among Xs if X has an extra
feature compared to Y.
Find the O
Q
Q
Q
Q
Q Q
Q
Q Q
QQ Q
O
Q
Q Q
Q
Q
Q Q
Find the Q
O
O O
O
O
O
OO Q O
O OOO
O O O
O OO
Illusory Conjunctions
When attention cannot be used to bind features together
because displays are too fast, then features free-float
independently, and may incorrectly recombine with each other.
Structural Analysis
Represent parts, and relations between parts
Geon theory
A fixed number of primitive geometric components
Composed in different arrangements to create all objects
Evidence for geons
Object recognition is hard if object cannot be analyzed into
geons
Combining geons to create objects
36 geons in all
Relations between parts is important
Recognition is easier when geons can be recovered
Practice does not improve Display Size X RT
slope for color-form conjunctive task
People do not form colorform units over time
Similarity of Targets to Distractors is not the
critical factor for determining search ease
What matters is
whether a simple
feature can be used
to detect a target
Well-learned
conjunctions are
still recognized by
combining their
features.
We do not form an
R unit?
Shiffrin & Lighfoot (1997)
Units made of simple line segments
can be created with practice
Unitization of complex forms (Goldstone, 2000)
Category 1
Category 2
ABCDZ
ABCDE
ABCYE
All task: Need to pay attention
to A, B, C, D, and E
One task: Only need to attend
one of the five segments
ABXDE
AWCDE
VBCDE
Responses to a conjunctive target with 20 hours of training
All
3,500
All
3,500
One
2,500
One
RT
2,500
1,500
1,500
500
500
J.W.
S.P.
All
3,500
All
3,500
One
2,500
One
RT
2,500
1,500
1,500
500
500
P.T.
C.E .
Blocks
Blocks
Holism revisited - Holistic sameness detector?
Task 1: Are these two grids cell-for-cell identical?
Task 2: Do these grids have any cells in common?
No
Yes
"Yes" judgments for Task 1 are faster than "Yes"
judgments for Task 2, even though, logically (in one
sense), Task 1 requires many Task 2-type judgments.
Expertise creates new perceptual units (Gauthier et al, 1999)
Fusiform gyrus (IT) activity due to novel stimuli (Gauthier et al, 1999)
Differences in object recognition due to expertise occur within 170
msec of stimulus onset (Curran & Tanaka, 2001)
Holism as interactions between parts (Farah, 1992)
Which is Larrys nose? (Part judgment)
Part in whole judgment
is much easier than part
judgment for faces
Which is Larrys face? (Part in whole judgment)
Faces are holistically
perceived
Tanaka & Farah (1993)
Feature Search for Race
The influence of life-long experience on face perception
Feature search task performance is more efficient with other race
targets than same race targets (Levin, 2000)
Conclusions
There is strong evidence for an analytic account of
pattern recognition
Parallel detection of features
Combining features together takes time and attention
Early perception of an object is as a bag of features
Holistic perception also occurs
Perception of entire forms without decomposition
Unitization of well-learned forms
Context effects on perception of features
Open Question
When do analytic and holistic perceptual processes occur?
What does it mean to be a psychological feature?
