A Model of Object Permanence

Psych 419/719

March 6, 2001

What is the Concept ofObject Permanence?

• The realization that an object still exists even if it is out of sight.

• Exploring when children seem to be able to use this concept tells us about their general cognitive development.

Early Usage..

• Infants at a very early age (3.5 months) seem to understand object permanence.

• Measured by gaze: infants look longer at “impossible” events; ones that violate the principle of object permanence– Impossible Event: Put a brick on toy train tracks.

Hide tracks and brick behind screen. Train goes along tracks and reappears at other end. Seems to go though the brick.

Reaching Tasks

• Even though children at age 3.5 months seem to be able to direct gaze based on object permanence, can’t always retrieve hidden objects.

• Not until about 8 months old can they retrieve hidden objects.

It’s Not Just Visual Input..

• Children younger than 8 months can retrieve objects that are in the dark, but not occluded by something.

• Implies that they don’t have to see the object to know where it is

• .. But they’re thrown by it being hidden by something.

One Standard Account

• Means-ends abilities: The ability to get at object X by manipulating object Y– Say, object X is under a blanket, or a pail, or

behind something.– Need to move that thing in order to get at object X.

• Claim: Children learn about object permanence around 3.5 months, but don’t learn means-ends abilities until 8 months.

Visually...Concept of Object PermanenceMeans-ends Abilities

3.5 Months 8 Months

“Kno

wle

dge”

The A~B Error

• When children can retrieve object, they still can make errors:– Hide object under pail to the left of child. Child

retrieves it. Repeat several times.

– Then, hide object under pail to the right of the child.

– Sometimes, child reaches to the left, out of habit

– But often their gaze is to the right

An Account of this Error

• Failure to inhibit an overlearned response.

• Problems:– Seems influenced by factors irrelevant to

inhibition, like cover on location A– A lot of these may just be random errors– Why would gaze and reaching have different

overlearned responses?

Characteristics of these Accounts

• Tend to view knowledge as all-or-nothing: you either have a skill or you don’t.

• When studying a given skill, then, evidence of success is evidence you have the skill

• So, failures must be attributed to something else that is outside of the skill you’re studying

• Implies you can ignore it...

The Competence / PerformanceDistinction

• Competence is defined as “having the skill”

• Performance is defined as your actual ability to do the task requiring the skill.

• These are theoretical constructs. They can be useful:– Throwing out noise from an experiment, like

coughs, slips of the tongue, etc.

The Danger of Taking TheseConstructs Too Seriously

• We always want to separate the true signal from noise.

• But: sometimes the noise isn’t really noise, but is diagnostic of your true “knowledge”.

• … Then we end up coming up with a lot of ancillary constructs to explain data (like “means-ends” or “inhibition”) that may be unnecessary

A Different Theory

• Knowledge is graded, not all-or-none

• Ability to apply knowledge is a function of the task. – Easy tasks can work with weak knowledge– Harder tasks require stronger knowledge or

representations

This Theory’s Account

• Young children have weaker representations of occluded objects

• These are strong enough to direct gaze– Gaze is an easy task

• Before 8 months of age, not strong enough to guide reaching

Occluded versus Darkness

• When an object is occluded, the child has visual evidence that it isn’t there. This works against weak representations of the object.

• When the object is in the dark, the child has no visual evidence that the object isn’t there.

• Recall: masking of stimuli from IAC model of word perception

Visually...

3.5 Months 8 Months

“Kno

wle

dge”

Enough to direct gaze

Enough to direct reach in dark

Enough to direct occluded reach

Evaluating The Theories

• “Principles”• Knowledge is all-or-

nothing• Failure to reach is

attributed to failure of means-ends ability

• “Graded Knowledge”• Representations build

up during development

• Failure to reach is a result of difficulty of task, not means-ends ability

An Empirical Test

• Previous experiments: confound between occluded-ness and means-ends.– Need to remove thing that hides object

• If failure to reach for occluded object is indeed a failure of means-ends ability, we should see no difference in performance for visible and occluded objects if both tasks require means-ends manipulation.

Their Experiments

• Devise a scenario where child needs to manipulate object X in order to get object Y– Whether Y is visible or not

• Expt 1: Child must pull blanket to get toy

• Expt 2: Child pushes button to cause toy to slide towards him.

Their Results

• Even with means-ends requirements equated, children 7 months old better at retrieving object that is visible than hidden– See Figure 3 (Expt 1), and Figure 5 (Expt 2)

• Conclusion: means-ends differences can’t account for why children have a harder time retrieving hidden objects.

A Model of the Graded-Knowledge Theory

• A visual representation encodes what is seen

• A context layer:– activated by its own state,

– and what is currently seen

– Predicts what will be seen next

context

Visual Rep

Encoding

Weights

Predictive

Weights

Recurrent Weights

The Visual Layer

• Seven units coding object position

• Seven more coding position of screen.

• If screen obscures object, object unit turns off.

What Training Prediction Does

• To predict accurately, network must know:– The current position of the obstruction– Direction of the obstruction, or whether it has

stopped moving– Where the object is in relation to the obstruction

• Must learn to store the location of the object!• Must learn to infer it will still be there

Evaluating the Network

• Plot sensitivity to occluded objects:

• Activity on prediction that ball will be there, minus activity on prediction when ball was not present

• Plot as a function of duration of occlusion (Fig 9)

Training

Sen

siti

vity

t3

t5

t7

What This Means

• The network learns, simply through predicting what it will “see” next, that objects that become hidden will be visible when the occlusion is removed.

• The ability of the network to do this develops gradually.

Simulating Reaching

• Train network to perform two tasks:– Predict next scene

– Reach for object

• Delay training on reaching task

• Also, use lower learning rate

context

Visual Rep

Encoding

Weights

Reach Units

Results

Training

Sen

siti

vity

Look 3

Performance on Looking

task (same as before)

Reach 3

Performance on reaching

for occluded object

Reach Vis

Performance on reaching

for visible object

What this Shows

• If feedback or capacity for reaching is reduced relative to that of gazing, then gaze develops more quickly than reaching

• This means that there is a point in development where infants will:– Correctly gaze at occluded objects– Correctly reach for visible objects– Not correctly reach for occluded objects

Interesting Caveat

• In the simulations, error was not propagated back from reach units to context units

• Hence, representations that develop in context units not sensitive to reaching task

• What might happen if they were?

For Next Class (Next Tues)

• Read PDP1, Chapter 5, “Feature discovery by competitive learning”

• Optional reading: handout

• Remember: Homework 3 due next Tues

• Remember: Project proposals due March 15