Page 1

Pattern Recognition

I. What is pattern recognition? II. Template Models III. Feature Models IV. Top-Down & Bottom-Up processing V. Neural Network Models VI. Prototype Models VII. Facial Recognition VI. Conclusions

I. What is Pattern Recognition

A. Definition: A process of identifying a

stimulus. Recognizing a correspondence between a stimulus and information in permanent (LTS) memory.

I. What is Pattern Recognition B. In the context of the Atkinson and Shiffrin Model Input Sensory Short- Long-

Store Term Term Store Store

Control Processes rehearsal coding retrieval strategies Response Output

Page 2

I. What is Pattern Recognition

C. This process is often accomplished with incomplete or ambiguous information.

D. Many variations on a pattern may be recognized as the same “object” or class of objects.

Page 3

Turing test (used by Yahoo, Hotmail, and ebay)

F. Pattern recognition that is difficult for machines is easy for people.

fi yuo cna raed tihs, yuo hvae a sgtrane mnid too.

I cdnuolt blveiee taht I cluod aulaclty uesdnatnrd waht I was rdanieg. The phaonmneal pweor of the hmuan mind! Aoccdrnig to a rscheearch at Cmabrigde Uinervtisy, it dseno’t mtaetr in waht oerdr the ltteres in a wrod are, the olny iproamtnt tihng is taht the frsit and lsat ltteer be in the rghit pclae.

The rset can be a taotl mses and you can sitll raed it whotuit a pboerlm.

Tihs is bcuseae the huamn mnid deos not raed ervey lteter by istlef, but the wrod as a wlohe. Azanmig huh?

[This demonstration is food for thought. The psychological principles it espouses are only partly correct. See Reicher (1969)]

II. Template Model

A. Basic Assumptions 1) Memory representation is a holistic

unanalyzed entity (a template). 2) An input pattern is compared to the stored

representation. 3) Identity is determined by selection of the

template with the greatest amount of overlap.

Page 4

II. Template Model

B. Schematic of a Template System Stimulus

Light Source Templates

Brightness Detector

A

C. Template systems in action

II. Template Model (cont)

Template Model (cont)

D. Problems with template models 1. Intolerance to deviations 2. Large number of templates required 3. Cannot support similarity-difference judgments

Page 5

III. Feature Theories

A. Basic Assumptions

1. The stored representation is a description of past inputs in terms of list of attributes or features.

2. Inputs are broken down into a small list of constituent features.

3. Identity is determined by selecting the feature list most similar to the input.

III. Feature Theories B. Schematic of a Feature Model

Stimulus

III. Feature Theories (cont) C. Supporting Evidence

1. Hubel & Wiesel (1962): Recorded electrical activity in the visual cortex of the cat.

Page 6

Hubel & Wiesel (1962) Results: specific cells respond to specific visual features.

III. Feature Theories (cont)

B. Supporting Evidence (cont) 2. Letter recognition times Gibson, Shapiro, & Yonas (1968)

Step 1: Analyze letters in terms of a small set of features.

Step 2: Give subjects a reaction test two determine if two letters are the

same or different. e.g. G vs.. W RT = 458 msec P vs.. R RT = 571 msec

Step 3: Compare the clustering of letters in the reaction time task to the similarities in features.

Step 1: Feature Analysis of Letters

Page 7

Step 2: Letter Groupings based on RT

III. Feature Theories (cont) D. Criticisms of Feature Theories

1. Importance of Context

2. Importance of Arrangement

IV. Top-Down vs. Bottom-up Processing

comprehension phrase processing word processing letter processing feature processing

Top Down (conceptually driven)

Bottom Up (data driven)

Page 8

IV. Top-Down vs. Bottom-up Processing

In Control of Attention (Bushman & Miller, 2007) implanted electrodes in monkeys

the monkeys were trained to search for a target in a visual display the researchers measured reaction time and recorded firing rates in parietal cortex (25 electrodes) (visual-sensory information) and the prefrontal cortex (25 electrodes).

IV. Top-Down vs. Bottom-up Processing

Bushman & Miller (continued) Bottom up: visual pop-out Sensory neurons (parietal) responded first

Page 9

IV. Top-Down vs. Bottom-up Processing

Bushman & Miller (continued)

Top down (visual search) prefrontal cortex responded first

Conclusion: Button up processing signals arise from the

sensory cortex. Top down processing signals begin in the

frontal cortex.

IV. Top-Down vs. Bottom-up Processing

V. Neural Network Model of Pattern Recognition

Feature Analysis

Word Analysis

Letter Analysis

Visual Input

A. Interactive Activation Model (McClelland & Rumelhart, 1981) Incorporates top-down processing from the word level to the letter level. Excitatory connections: Inhibitory connections:

Page 10

A T H C

CAT THE CHAIR

E

Input

Feature Level

Letter Level

Word Level

Excitatory connections: Inhibitory connections:

Simplified view of the Network of Connections

More Complete view of the Network of Connections:

B. Supporting Evidence:

The word/letter effect Reicher (1969)

Stimulus Example Test Percent Correct letter h h/t series csah csah/csat word cash cash/cast

78

76 89

Page 11

VI. Prototype Theory

A. Basic Assumptions 1. The stored representation is a Prototype: an

abstraction of the typical or best example of an object. examples: chairs, cars, and trucks

2. Inputs are broken down into feature lists. 3. Recognition is process of comparing the

features of the input to the features of prototypes, and selecting the best fit.

B. Evidence for Prototype Theory Solso & McCarthy (1981) face recognition

VI. Prototype Theory (cont.)

Prototype 100%

75%

50%

25%

0%

100 75 50 25-5

-4

-3

-2

-1

0

1

2

34

5

100 75 50 25

Old ItemsNew Items

VI. Prototype Theory (cont.) Solso & McCarthy (1981): results

Con

fiden

ce

Old

New

Percent Overlap with Prototype

Page 12

VI. Prototype Theory (cont.)

C. Prototype Theory and attractiveness 1) “goodness” of category membership can be defined with respect to the prototype.

2) “good” category members may be seen as more attractive, or desirable, than “poor” category membership

C. Prototype Theory and attractiveness (cont.)

Example: “attractive faces are average” (Langlois & Roggman, 1990) Stimulus set:

individual faces composite faces containing 2 - 32 faces.

Examples of composite faces: Number in composite 4 8 16 32

Page 13

Rated attractiveness

Number of faces average rating 1 2.51 2 2.87 4 2.84 8 3.03 16 3.06 32 3.25

VII. Facial Recognition:

Why Barack Obama is Black (Halberstadt et al, 2011)

Hypodescent: association of mixed race individuals as belonging to the minority race.

Hypothesis: individuals learn to minority

groups later than majority groups, so they learn to focus attention on features that distinguish the groups.

Increased attention to distinctive (distinguishing) features leads to over-classification in the “new” group.

Why Barack Obama is Black (Halberstadt et al, 2011)

Evidence: Experiment 1

Participants: ½ Caucasians (New Zealanders)

½ of Chinese decent (raised in China or Asian Pacific regions).

Individuals performed a speeded classification of faces that were morphed blends of Chinese and Caucasian faces:

Page 14

Why Barack Obama is Black (Halberstadt et al, 2011)

Experiment 2

Participants: 75% Caucasian, 25 % other Procedure: participants learned to classify faces into different (arbitrary) groups. “majority faces” classified 9 times “minority faces” classified 3 times

Why Barack Obama is Black (Halberstadt et al, 2011)

Results Experiment 1 Percent of ambiguous faces rated as Chinese: Chinese Participants: 44 %

Caucasian Participants: 49 % Experiment 2

Percent of ambiguous faces rated as B’s A faces “majority”: 40 % B faces “majority”: 36 %

Conclusions:

Biracial classifications are based on learning history. Distinctive racial features receive greater attention if they are learned later in life.

Why Barack Obama is Black (Halberstadt et al, 2011)

Conclusions: Biracial classifications are based on learning history. Distinctive racial features receive greater attention if they are learned later in life.

Page 15

VII. Facial Recognition:

A special problem for theories of pattern recognition:

A. Different set of rules? (Example: object vs.

facial recognition). Yin (1970), and Rock (1974) demonstrated that

facial recognition is more easily impaired by inversion than is object recognition.

Who is this?

Page 16

A B

Page 17

A B

VII. Facial Recognition (cont)

B. Different Neurological Structures? Dissociation between loss of object recognition

(visual agnosia) and face recognition in stroke victims.

(e.g., Msocovithc, Winocur, & Behrman, 1997)

VI. Conclusions on Pattern Recognition

A. Template and Feature Models are inadequate

B. Context and top-down processing are very

important C. Neural Networks can explain top down

processes. D. Important role of prototypes E. Challenge of explaining facial recognition