The Influence of Prior Knowledge on Recall for Height

Jenny ShiPernille Hemmer

Mark Steyvers

Department of Cognitive ScienceUniversity of California, Irvine

1

Introduction

Memory is not an exact replica of an event, but rather more of a noisy representation. We fill in the gaps to make sense of the memory. This “fill-in” is influenced by our expectations and prior knowledge.

People usually depend on this prior knowledge to draw conclusions, recall certain events, make decisions, solve problems

• Expectations, stereotypes, common sense

...but does prior knowledge help or harm our recall from memory?

2

Prior knowledge harms

Much of previous literature has focused on negative impacts of expectations.

Participants falsely recalled books in an office setting, possibly because most offices are expected to contain books.

(Brewer & Treyens, 1981)

3

Prior knowledge harms

Semantic word lists elicit false recall of words.

(Roediger & McDermott, 1995)

These studies show that expectations are detrimental to recall from memory because they induce false memories.

4

BEDSNOOZEDREAMWAKERESTDOZEYAWN

SLEEP

Directions:

Recall words from the list

Prior knowledge helps

Memory task with vegetables, fruits, and artificial shapes.

Reconstruction from memory was better for objects that subjects had pre-experimental prior knowledge for compared to objects they had no prior knowledge for.

(Hemmer & Steyvers, 2008)

5

Current Study

Extension of Hemmer & Steyvers (2008)

Efficiency in analysis

• Hemmer & Steyvers used the prior knowledge of others as their true value. The current study uses the actual true value and is independent from subjectivity.

Moving towards naturalistic stimuli

• While the sizes of fruits and vegetables can vary across culture and time, human height stays relatively constant. The average height of an American male isn’t much different from an average European male.

6

Why use height as prior knowledge?

Prominent difference between the mean height of males and females

• Different distributions and means between two genders

Objective scale of measurement

Consistency in nature

• Stereotype is drawn from real-world distributions

• Accurate and universal across societies

The concept is deep-seated

• In a study, subjects still used this prior knowledge even though they were told that gender and height had no correlation (Nelson, Biernat, & Manis, 1990).

7

Absolute Height

Relative height of males

Relative heightof females

Overall distribution of heights

A

B

We predict…

A. Regression to mean• People are estimated towards the mean.

• People shorter than the mean are overestimated while people taller than the mean are underestimated.

B. Hierarchical effect• If prior knowledge is more fine-

grained, people are estimated towards the mean of their specific category.

• If tall female and short male presented at same height, the female will be underestimated while the male will be overestimated.

8

Stimuli Production

56 58 60 62 64 66 68 70 72 74 760

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Height

CDF

female population

male populationfemale sample

male sample

Height

CDF

9

• Photographed total of 200 males and females with digital camera

• Height recorded in inches

• Stimuli height corresponded to true population

Height (inches)

Stimuli Production

Unmasked stimuli – Retained all aspects of the figure, essentially revealing prior knowledge of gender

• 48 images were selected for this category: 24 ♂, 24 ♀

Masked stimuli – Retained only height and human figure info, restricting prior knowledge of gender

• 24 stimuli selected

• Silhouette of unmasked stimuli

• Gender ambiguity ensured through a ratings panel evaluation

Unmasked Male Unmasked Female

10Masked

Experiments

Two separate experiments (Between-subjects)

Prior knowledge task

• Assessment how subjects’ prior knowledge affect height estimations

• Participants asked to judge how tall a person is based on whether prior knowledge of gender is restricted or revealed.

Memory task

• Assessment of how prior knowledge affects memory

• Continuous recall paradigm: Study and test phases are randomly intertwined

11

Prior knowledge task

12

What do you think is the true height of this person, compared to the door on the left?

OK

Reference image of a door

Experimental Design for Memory Task

2 blocks with the memory task

• Block 1: 48 unmasked stimuli (24 male, 24 female)

• Block 2: 24 masked stimuli

Each stimuli was used for one trial (72 total)• Counterbalanced between and within blocks

13

Memory Study – Study phase

14

Memory Study – Test phase

What was the height of this person, compared to the door on the left, when you saw them at study? If you

are not sure, make a best guess.

OK

15

Memory Task: General Results

A linear regression model fit to the data• Slope (regression to the mean)

• Intercept (hierarchical effect)

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Rec

onst

ruct

ed -

Stu

died

Female

Male

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Rec

onst

ruct

ed -

Stu

died

Masked

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Rec

onst

ruct

ed -

Stu

died

Female

Male

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Rec

onst

ruct

ed -

Stu

died

Masked

16Original Study Size

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Reconstr

ucte

d -

Stu

die

d

Female

Male

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Reconstr

ucte

d -

Stu

die

d

Masked

Memory Task: Unmasked Results

Biased towards the mean height• Negative slopes indicate that

short people are overestimated and tall people underestimated

Each gender is biased towards their own specific height distribution

• Females estimated to be shorter than males

Directional biases based on gender info• When presented in same height,

females estimated shorter, while males estimated taller

17

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Reconstr

ucte

d -

Stu

die

d

Female

Male

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Reconstr

ucte

d -

Stu

die

d

Masked

Memory Task: Unmasked Results

Biased towards the mean height • Negative slopes indicate that

short people are overestimated and tall people underestimated

Each gender is biased towards their own specific height distribution

• Females estimated to be shorter than males

Directional biases based on gender info• When presented in same height,

females estimated shorter, while males estimated taller

18

Memory Task: Masked Results

Bias corresponds to the general distribution of human height

Our shortest stimuli was recalled with near perfect accuracy

• Possible outlier effect

• Could be attributed to a cue other than gender that’s inferring a shorter stature

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Reconstr

ucte

d -

Stu

die

d

Female

Male

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Reconstr

ucte

d -

Stu

die

d

Masked

19

Memory Task: Masked Results

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Reconstr

ucte

d -

Stu

die

d

Female

Male

140 160 180 200-30

-20

-10

0

10

20

30

Original Study Size

Reconstr

ucte

d -

Stu

die

d

Masked

Our shortest stimuli was recalled as short on average.

20

Task Type Results

Recall error was greater for masked stimuli than unmasked.

• Restriction of gender information introduces greater bias.

Less error for Memory task compared to the Prior knowledge task.

Even when participants have not studied an image, performance is better for unmasked compared to masked stimuli.

• Prior knowledge is helpful

Female Male Masked0

2

4

6

8

10

12

14

16

18

Reco

nstru

ction

Erro

r (M

ean

Abso

lute

Divia

tion)

Memory

Prior Knowledge

Baseline

Unmasked Female Unmasked Male Masked

Conclusion

By using naturalistic stimuli, we were able to obtain results that suggest having prior knowledge improves reconstruction from memory.

Influence of prior knowledge is hierarchical.

• General prior knowledge will create bias towards mean of general distribution.

• Access to finer-grained prior knowledge will differentially affect the memory for two people (male and female) originally presented at the same height. They will be biased towards the mean of their own specific distribution.

22

Acknowledgments

Thank you to…

Pernille Hemmer

Mark Steyvers

Michael D. Lee

Memory and Decision-Making Lab, UCI

23

Extra slides

24

Gender Height Distributions

58 60 62 64 66 68 70 72 74 760

0.05

0.1

0.15

0.2

0.25

Height in Inches

Male

Female

25

Stimuli Production

STIMULI MEAN

TRUE MEAN

MALES 68.9 69.4†

FEMALES 63.9 63.8†

† = Figures obtained from McDowell et al. 2008. Figures reflect the U.S. population over age 20.

Gender mean heightOur data:

Mean height Female: 63.9 inches (SE 0.24)

Mean Height Male: 68.9 inches (SE 0.36)

In US population (20-29 years old):

Mean height Female: 64.5 inches

Mean height Male: 70 inches

(US census: http://www.allcountries.org/uscensus/)

In US population over 20:

Mean height Female: 63.8 inches (SE 0.06)

Mean height Male: 69.4 inches (SE 0.07)

(CDC: http://www.cdc.gov/nchs/data/nhsr/nhsr010) 27

Norming masked

100 120 140 160 180 200 220

100

120

140

160

180

200

220

estimated height

Tru

e he

ight

Female

Male

28

Correlation between estimated and true height of masked females and males. Both means for the genders are the same (about 155 cm).

Norming unmasked

100 120 140 160 180 200 220

100

120

140

160

180

200

220

estimated height

Tru

e he

ight

Female

Male

29

Correlation between estimated and true height of unmasked females and males. Means for the genders are the different (female 140 cm; male 150 cm).

Norming unmasked (female)

100 120 140 160 180 200 220

100

120

140

160

180

200

220

estimated height

Tru

e he

ight

Female

linearMale

30

Norming unmasked (male)

100 120 140 160 180 200 220

100

120

140

160

180

200

220

estimated height

Tru

e he

ight

Female

Male linear

31

Potential extensions...

Labeling• Show masked but also include a label of either “Male”,

“Female” or “Unknown”

• Can control the specific prior knowledge subjects receive.

• Differentially recall height based on the gender info provided

Outlier Effect• We didn't see one in the current study.

• Main problem was the stimuli.

• We need people with extreme heights (esp. males)

32