Psychology 85-419/719 January 30, 2001

Processing and Constraint Satisfaction:

Psychological ImplicationsThe Interactive-Activation (IA)

Model of Word Recognition

Psychology 85-419/719January 30, 2001

Background: Fodor’s Definitionof Modularity

• Input systems are composed of distinct processing modules

• Modules have certain properties:– Partial results not shared between modules.

Communication is all-or-nothing.– Information is encapsulated in a module; only

results are shared. Access to global data is limited.

• Contrast with view we’ve seen so far...

More Background: Models and Demonstrations

• A model is our best attempt at simulating a system. We think it’s basically true (to a certain level of approximation, anyway).

• A demonstration is a simulation that we know isn’t right, but that demonstrates a useful point; say, about computational principles.

• Which one is the IA model?

Letter Perception In Context:The Phenomena

DOG#### BNR####

O or U? N or M?

The Empirical Findings...

• Subjects are more accurate in identifying letters in briefly presented stimuli if the letter was in a word (as opposed to random letters, or individual letters)

• Nonwords that are pronounceable (e.g., MAVE) show advantage over non-pronounceable strings (e.g., MVAE)

Assumptions of the Model• Fodor’s wrong.

– Processing is interactive, parallel, with partial results feeding different representational areas.

• There are (at least) 3 levels of analysis: features, letters, and words.

• The levels inhibit or excite each other depending on whether they are consistent with each other.

• Context effects can emerge from interactions between levels of representation

The Overall Model

Word Level

Letter Level

Feature Level

Visual Input

“Context”

Acoustic Level

Phoneme Level

Acoustic Input

ImplementedModel

Spelling Speech

Representations of Visual Features

16 features, each corresponding to a linesegment.

4 slots, one for each letter.

Levels of Representation

Word Level:Inhibitoryconnections

cat dog lake

cLetter level:Inhibitory andexcitatory connections

a

Feature Level(still more connections)

Pre-Set Weights

• Negative, inhibitory weights between word nodes. All same value.

• Positive or negative weights between letter nodes and word nodes, and between feature nodes and letter nodes. Same values for all weights.

• Biases on word nodes a function of word’s frequency.

Processing

• Generally, the same formulation that we’ve been working with:– Network of weights, activities for units over

time.

• … with additional mathematics to simulate a forced-choice response

The Running Average(Eq. 5)

t rxt

ii dxexata )()()(

rxte )(

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 2 4 6 8 10 12 14 16 18 20

Time

Weighting

r=1

r=0.05

r=0.01

In simulations, r=0.05

Response Probability(Eq. 6 & 7)

)()( taui

iets Strength of option i is:

Probability of Response for i is:

jj

ii ts

tstRp

)(

)(),(

In simulations, u = 10

Model Behavior: Degraded Input

work

word

wear

Letter Activations Too:

k

a

r

The Word Preference Effect

• When stimuli is masked, letters embedded in words are perceived more accurately than letters standing alone

“e” in “read”

“e” alone

Mask applied

Probabilities...P

roba

bili

ty

0

1

“e” in “read”

“e” alone

Subjects: 80% for word 65% alone

Model: 81% for word 66% alone

Difference Between Maskedand Degraded Stimuli

• When stimuli is masked, there is actual information actively disrupting the visual system– By hypothesis, this actively turns off the letter

representations

• In contrast, when stimuli is simply degraded, there is still some activity in letter units. It’s noisy, but not obliterated.

Simulating Masking andDegraded Stimuli

• Masking: Present stimuli in reliable fashion for period of time. Then, activate all segments (corresponds to mask)– Result: suppress all letter nodes

• Degraded stimuli: Present stimuli where features have a probability of being detected.

Interactions...

Empirical Data

80 78

6573

0

20

40

60

80

100

Masked Degraded

In WordAlone

Simulation Results

81 7866 68

0

20

40

60

80

100

Masked Degraded

In WordAlone

Why?

• In masked condition, letters lose all visual excitation.

• All activity, then, is a result of top down influences. For words, this is much larger than for single letters.

• In degraded stimuli, there is still some visual information. So single letters not so reliant on top down information.

Letter Perception in Nonwords

• Nonwords that look like words (e.g., MAVE) show letter advantage over letters in isolation too.

• IA model account:– Even though MAVE may not “win” with any

word nodes, it overlaps with enough word nodes (GAVE, SAVE, HAVE) for the letters to get some top-down support

Neighbors, Friends and Enemies

• A neighbor of a word is one that differs only by one letter

• A letter (e.g., M) in a spelling pattern like MAVE has friends; words that are neighbors and have an M in the 1st position (MOVE, MAKE, MADE)

• There are also enemies. Words that are neighbors but don’t have an M in the 1st position (HAVE, GAVE, SAVE)

The “Rich Get Richer” Effect

Have (high frequency)

Gave (medium frequency)

Save (low frequency)

The “Gang Effect”

Save. Part of large gang

Male. Also, large gang

Move. Member of smaller gang

Other Phenomena?

Word Level

Letter Level

Feature Level

Visual Input

Acoustic Level

Phoneme Level

Acoustic Input

Lesch & Pollatsek ‘93:TOWED primes FROG

Semantic Priming:TOAD primes FROG

Impairments?

Two-hop: STRIPESprimes LION

The “Slot Problem”

Wrapping Up Section I: Constraint Satisfaction

• Complex patterns of behavior arise from “simple” interactions between processing units

• Weights encode knowledge about relationships between atomic facts, propositions, perceptions

• Networks are dynamic; representations evolve over time

Next Section: Simple Learning

• For next class: read handout (from Handbook, Chapter 4, pages 83-89; see web page)

• Homework 1 due (but two day grace period)

• Next homework handed out. Due Feb 15.