Learning and Transfer in Context Robert Goldstone Ji Son Indiana University Department of Psychological and Brain Science Program in Cognitive Science.

Learning and Transfer in Context

Robert GoldstoneJi Son

Indiana UniversityDepartment of Psychological and Brain Science

Program in Cognitive Science

How do we learn in a way that is general and transportable?

• Through strategic decontextualizations

• By learning contextualized, grounded system

Arguments for decontextualized scenarios• Transfer depends on understandings that are not tied to their context

• Better transfer from simplified to rich forms than vice versa (Bassok & Holyoak, 1989; Son, Smith, & Goldstone, 2008)

• Better transfer from idealized, relative to concrete, forms to an isomorphic situation (Goldstone & Sakamoto, 2003; Goldstone & Son, 2005)

• Better understanding of the formal properties of a system from simple symbols than rich instances (Kaminsky, Sloutsky, & Heckler, 2008; Sloutsky, Kaminski, & Heckler, 2005)

• Competition between symbolic and physical interpretations of objects (DeLoache, 1991; Uttal, Liu, & DeLoache, 1999)

Info Reset Adjust Quiz

Paint Erase Place Move

Clear

Graph Updating

Number of ants Pen Size

Closest Rate Not Closest Rate

Time

IdealizedConcrete

Son, Smith, & Goldstone (2008)

Goldstone & Sakamoto (2003)

Kaminsky, Sloutsky, & Heckler (2008)

Uttal, Liu, & DeLoache (2006)

IdealizedConcrete

Arguments for contextualization• Reasoning is facilitated when the domain is concrete rather than abstract

(Wason & Shapiro, 1971, Johnson-Laird, Legrenzi, & Legrenzi, 1972)

• Knowledge is grounded in perception (Barsalou, 1998), action (Glenberg, 1997), or visuo-spatial models (Lakoff & Nuñez, 2000).

• People often perform tasks better in rich, real-world contexts than in laboratories or classrooms (Lave, 1988; Nunes, Schliemann, & Carraher, 1993)

• Learning that is grounded in authentic learning experiences is often more meaningful to students (Barab & Roth, 2006; Ladson-Billings, 1995)

Kinds of Contextualizations

Context Contextualized DecontextualizedPerceptual Visuo-spatial

grounding, realistic and rich

Amodal, symbolic, idealized

Experiential Active exploration Direct instruction

Textual voice First or second person (I, you)

Third person (she)

Narrative Specific, rich General, sparse

Problems Real-world, authentic

Invented, Idealized

Cultural Culturally meaningful

Outside of one’s culture

Kinds of Contextualizations

Context Contextualized DecontextualizedPerceptual Visuo-spatial

grounding, realistic and rich

Amodal, symbolic, idealized

ExperientialExperiment 1

Active exploration Direct instruction

Textual voiceExperiment 2

First or second person (I, you)

Third person (she)

NarrativeExperiment 3

Specific, rich General, sparse

Problems Real-world, authentic

Invented, Idealized, arbitrary

Cultural Culturally meaningful

Outside of one’s culture

Experiment 1: Experiential contextualizationSon & Goldstone (in press, Cognition & Instruction)

• Experience-based learning– Constructivism (Savery & Duffy, 1994)– Problem-based learning (Grabinger & Dunlap, 1995)– Active control by learner (Schwartz & Bransford, 1998)

• Direct instruction– More children learn about experimental control from direct instruction than discovery learning

(Klahr & Nigam, 2004)– Cognitive load (Kirschner, Sweller, & Clark, 2006)– Possible inefficiences of problem-based learning (Albanese & Mitchell, 1993)

Signal Detection Theory (SDT)• Decision making under uncertainty

– Target is either present or absent– Observers judge presence based on signal strength in noise– Critical notions: evidence, decision, criterion, outcomes (false alarms, correct rejects, hits, misses)– Useful because of wide applicability

• All subjects receive a graphical and textual tutorial– A doctor diagnosing patients as having leukemia based on distortion level of cells– Half of subjects receive categorization experience - see distorted cells and categorize patients, with feedback– Training quiz

• Transfer to second SDT situation– Farmers diagnosing melons as ripe or not, based upon the imperfect predictor of weights– Transfer Quiz

Pure Distorted

Most healthy people have low cell distortion.

Pure Distorted

Most sick people have high cell distortion

S

S

S

S S S S

SS

S

SH

H

H

H

H

HH

The doctor decides these people are

“healthy”…

…and these people are

“sick”.

S

S

S

S S SS

S

S

S

S

H

H

H

H

H

HH

Later on, he found out who was actually sick and who was actually healthy:

S

S

S

S S

SSS

S

S

S

H

H

H

H

H

HH

Later on, he found out who was actually sick and who was actually healthy:

S

S SSS

S

S

S

H

S

S S

H

H

H

H

HH

“healthy” / sick

“sick” / sick

“healthy” / healthy

“sick” / healthy

S

S SSS

S

S

S

…BUT also decides “sick” when they are

actually healthy.S

S S

H

HH

After… he does not decide

“healthy” when they are actually sick…

S

S

S

S

“healthy” / sick

“sick” / sick

“healthy” / healthy

“sick” / healthy

The number of actually healthy and unhealthy people are the same two months in a row. However, in the second month, the doctor is diagnosing more patients as healthy when they are actually healthy and more people as healthy when they are actually unhealthy. What must have changed in the second month?

A)The doctor must be diagnosing people with more distorted cells as healthy.

B)The doctor must be diagnosing people with less distorted cells as healthy.

C)The doctor must be diagnosing more people who are actually healthy as unhealthy.

D)The doctor must have become better at diagnosing healthy people.

Doctor Quiz

Does active experience help?

Doctor diagnosis experience + tutorial

Transfer story about a French town trying to export sweet melons.

Melon quiz

Only tutorial

Only for Experiencing Participants

Experience hurts performance (particularly transfer)

• Playing with a model interferes with treating it symbolically (DeLoache, 2005)

• Students interpret the training scenario too narrowly in terms of doctors rather than more generally in terms of signal detection theory

Experiment 1 Results

Which of the following decision strategies will ensure that the doctor maximizes the number of actually healthy people he diagnoses as healthy?

A. diagnose everyone as healthy.B. look more carefully at the cell distortion levels before his diagnosis.C. examine the ratio of sick patients with distorted cells to sick patients with pure cells before his diagnosis.D. examine the ratio of patients with distorted cells to patients with pure cells before his diagnosis.

6

17

4

0

10

24

2

14

0

5

10

15

20

25

A B C D

how many peoplectrl

exp

Correct but unintuitive

Experiment 2: Textual voice

• Contextualization through grammatical voice of narrative• The second-person pronoun (“you”) has been recommended for text

books: “The use of the second person pronoun you also connects the reader to the mathematics because it speaks to the reader directly” (Herbel-Eisenman & Wagner, 2005; Moreno & Mayer, 2004)

• Third-person voice (“he”) fosters position of reader as detached, omniscient observer (Duchan, Brewer, & Hewitt, 1995; Graesser et al., 1996)

• Does placing participants in the story by use of the second-person pronoun help or hurt comprehension and transfer of principles?

Imagine a doctor who looks at blood samples to check if his patients have leukemia, a

cancer of the bone marrow.

Since bone marrow produces blood cells, the doctor can look for distorted blood cells to

help him diagnose his patients.

Imagine that you are a doctor who looks at blood samples to check if your patients have

leukemia, a cancer of the bone marrow.

Since bone marrow produces blood cells, you can look for distorted blood cells to diagnose

your patients.

Experience hurts performance (particularly transfer)

• Greater contextualization interferes with domain-general transfer of SDT principles

• Cost of maintaining counter-factual?

Experiment 2 Results

The towns sets a 1750 gm minimum weight but they do not know which melons are sweet or bitter. How many melons that weighed 1500 gm were rejected?

Third-person

Second-person

Only including sweet melonsIncluding all melons

Experiment 3: Narrative specificity

• Contextualization by presenting a specific, well-known scenario• Vivid information is better remembered than pallid information

(Nisbett & Ross, 1980)• Education researchers have recommended presenting materials with

well-known celebrities and culturally relevant contexts (Ladson-Billings, 1995)

• But salient specific details may interfere with extraction of abstract structure and transfer (Kaminski, Sloutsky, & Heckler, 2005, 2008)

• Initial scenario: either generic doctor or Derek Shepard™from the TV show “Gray’s Anatomy”

From C. Kea, “Infusing culturally responsive activities in your lesson plans”

Celebrity context hurts training performance, but not transfer

Experiment 3 Results

A strong difference between generic and specific conditions is only found for participants familiar with show

Generic Specific

Experiment 3 Results

Conclusions

• Distinguish types of contextualization• Increasing contextualization is not always a good thing• Tension between scaffolding understanding by a concrete

grounding and tying knowledge too tightly to a specific domain

• Potential dissociation between engagement and deep understanding

Approximately what percentage of all 1000 gram melons (1 kg) are sweet?

A. 10%B. 25%C. 33%D. 50% E. 66%

6

18

10

0

2

10 10

14

0

7

0

5

10

15

20

A B C D E

how many peoplectrl

exp

A. 10%B. 25%C. 33%D. 50% E. 66%

Experience people aren’t taking longer…

22.22221.85

0

5

10

15

20

25

1

time in minutes

ctrl exp

Controls are more semantic, experience people are more goal-biased

0.31

0.520.48

0.69

0

0.25

0.5

0.75

1

semantic answers goal-biased answers

percent correct in analogy

ctrl

exp

Sick <-> Bitter Sick <-> Sweet