Individual Differences and Dissociations in Category Learning (CL) Tasks Alan Pickering and Ian Tharp Department of Psychology [email protected].

Individual Differences and Dissociations in Category

Learning (CL) Tasks

Alan Pickering and Ian TharpDepartment of Psychology

[email protected]

Collaborators

Luke Smillie (University of Queensland)

Rozmin Halari (Institute of Psychiatry)

Lucy Schomberg, Debbie Benson, Fiona MacNab, and Wasima Ahmed

(St George’s Hospital Medical School)

Multiple Systems in CL?CL tasks may perhaps draw on 3 separate learning/memory systems:Explicit verbalisable rule system

prefrontal cortexProcedural (implicit) system

basal gangliaEpisodic memorisation (exemplar) system

medial temporal lobes (MTL)

Multiple Systems in TasksSystem Key

processDeployed in CL tasks with:

Rule-based Workingmemory

many exemplars; verbalisable rule;irrel. dimensions

Procedural Reinforce-ment

many exemplars; no rule; feedback;info integration

Episodic Exemplarstorage / retrieval

few exemplars; no rule

CURRENT APPROACH

• Uses individual differences (esp. personality trait scores), in healthy subjects, as a tool for exploring dissociations in category learning (CL) tasks

• Looking for a characteristic “individual differences signature” for each different type of CL task

Which Personality Traits?

• Extraversion-Introversion (EXT)Example measures: EPQ-E; Introvertive Anhedonia (IntAnh)

• Impulsive Antisocial Sensation Seeking: (IMPASS)Example measures: EPQ-P; Novelty Seeking; Sensation Seeking Scale

• Positive Schizotypy (SCHIZO)Example measures: Unusual Experiences

Example Questionnaire Items• IMPASS: Measure = EPQ-P (25 items)

-Have people said that you sometimes act too rashly?-Should people always respect the law?-Would you take drugs which may have strange or dangerous effects?

• Positive Schizotypy: Measure = Unusual Experiences (30 items)-I have felt that I have special, almost magical powers-Do you ever feel that your thoughts don’t belong to you-Sometimes my thoughts are as real as actual events in my life

Biological Basis of IMPASS: Sample Evidence

• Gray, Pickering & Gray (1994): SPET DA D2-binding in basal ganglia and EPQ-P

R2 = 0.5433

1.25

1.5

1.75

2

0 1 2 3 4 5 6 7 8 9

EPQ-P Score

DA

-D2

Bin

din

g I

nd

ex

LEFT

R2 = 0.5598

1.25

1.5

1.75

2

0 1 2 3 4 5 6 7 8 9

EPQ-P ScoreD

A-D

2 B

ind

ing

In

de

x

RIGHT

Category Learning & IMPASS: Previous Work By Others

• Ball & Zuckerman (1990): positive correlation between Sensation Seeking scores and learning of a concept formation task

• Task likely to be rule-based but had relatively few exemplars and employed feedback

Category Learning & Personality: Our Previous Work

1. IMPASS traits correlated positively with CL performance in 2 studies with Kruschke (1993) task

2. Results ambiguous: task could be solved by a simple rule requiring selective attention to 1 of 2 dimensions, but it had only 8 exemplars, and training involved feedback

Category Learning & IMPASS: Study 1 with Ahmed

• A rule-based task (Kruschke, 1993)• 2 stimulus dimensions: one predicts

category membership, other irrelevant• Only 8 exemplars• Training used verbal feedback• N=30 healthy male med. Students• Measured IMPASS using Novelty

Seeking Scale of Cloninger

Task: After Kruschke (1993)

Results: Effects of IMPASS

Learning Epoch

654321

Me

an

Nu

mb

er

of

Co

rre

ct

Re

sp

on

se

s (

/8)

8

7

6

5

4

3

Low

High

Category Learning & Personality: Study 2 (Benson/MacNab)

• Same task (Kruschke, 1993) but with two phases

• N=51 healthy med. students • Measured IMPASS (EPQ-P) and SCHIZO

(Unusual Experiences)

Regression Results

Regression Standardized Predicted Value

210-1-2-3

Nu

mb

er

Co

rre

ct

Ov

er

2 P

ha

se

s

100

90

80

70

60

50

R2=0.16; (IMPASS)=0.29; (SCHIZO)= -0.43

Interpretation & Conclusions: 1• IMPASS personality traits appear to be

reliably related to CL task performance, and evidence for positive schizotypy traits too

• Unclear which learning system(s) may have been predominant in the task used.

• Further studies with careful task comparisons needed

Category Learning & Personality: Our Previous Work

3. Double dissociation found using matched tasks: EXT was associated positively with task A but not B whereas reverse pattern was found for IMPASS.

4. Task A encouraged use of procedural learning system (reinforcement; no rule; info integ structure; probabilistic); whereas task B did not (paired-associate training without reinforcement)

Study 3 (with Halari)• Within-Ss design using 2 equivalent

probabilistic category learning tasks: Weather task and a Symptoms-Disease task

• Learning Regime (c/b across tasks, order)RF : enhanced reinforcement

£0.10 per correct responseinfo. integration structure

PA : paired-associate trainingmeant no reinforcement

• TestingCategorise each stimulus without reinforcement

Category Learning in Parkinson’s Disease

Weather task: Knowlton et al, 1996

Details• 40 healthy male participants, mostly

students• Personality Measures

EXT: Introvertive Anhedonia (IntAnh)

IMPASS: EPQ-PSCHIZO: Unusual Experinces

(UnEx)• Dependent Variable = Accuracy of

responses during test

Results: CorrelationsRF= reinforcement task scorePA= paired-associate task score

SCHIZO EXT IMPASS

RF UnEx IntAnh EPQ-P

PA 0.14 -0.01 -0.06 0.30*

RF 0.20 -0.35* 0.00

UnEx 0.02 0.34*

IntAnh 0.13

Interpretation & Conclusions: 2• Extraversion measures correlate with CL

task performance where procedural system involvement is encouraged.

• Fits with neurobiological models of extraversion.

• IMPASS traits also correlate with CL task performance where procedural system involvement is unlikely. But why?

Scanning to the rescue …?•Poldrack et al 2001’s fMRI study in

healthy volunteers with weather task using standard feedback (FB) vs. paired associate (PA) training

•Medial temporal lobe (MTL) activity higher for PA than for FB task

•Reverse was true for basal ganglia (caudate nucleus) activity

•Maybe IMPASS correlation with PA task is mediated by episodic memory

Category Learning & Personality: Our Previous Work

5. Showed that IMPASS measures correlated positively with episodic memory performance in 2 studies.

6. IMPASS measures also found to be correlated with behaviour on other tasks associated with hippocampal/MTL functioning (latent inhibtion; response to associative mismatch)

Study 4: IMPASS and Paired Associate Learning

• Pickering and Schomberg• Unrelated verbal paired-associates

(e.g. SOIL-MILE; SIDE-BRAVE) were used

• This is the quintessential explicit memory task sensitive to hippocampal lesions

• 40 healthy subjects (students)• Extraversion (Ext), IMPASS

(EPQ-P), and positive schizotypy (UnEx) were measured

ImpASS and Paired Associate Learning: II

• 12 word pairs (A-B) used• 3 study-test learning trials• Test= cued recall for B using A as cue • 1 unexpected 10-min delayed cued

recall test trial• DVs=Number correct on each test

(NC1, NC2, NC3, & NCD)• Measured IQ subtest performance for

each subject (WAIS-III Matrices)

IMPASS and Paired Associate Learning: Results

*denotes correlation with EPQ-P after partialling out IQ

EXT IMPASS

IMPASS*

SCHIZO IQ

NC1 0.17 0.43 0.40 0.15 0.26

NC2 -0.03 0.33 0.29 0.09 0.28

NC3 -0.10 0.33 0.29 0.08 0.27

NCD -0.04 0.37 0.35 -0.01 0.15

Interpretation & Conclusions: 3• IMPASS traits appear to be positively

associated with performance on hippocampal-sensitive, episodic memory tasks

• The repeated positive correlation between IMPASS traits and CL task performance may therefore be indicative of the involvement of episodic memory processes on those tasks

IMPASS and CL: Study 5• Study by Ian Tharp using matched

information integration (II) and rule-based (RB) tasks from the Ashby/Maddox stable

• Counterbalanced within–Ss design• 16 training exemplars with feedback,

trained to criterion• 82 healthy subjects (mostly students)

Ashby et al: RB Task• 1 dimension (background colour) determines category• Readily verbalisable rule

Cat B

Cat A

Ashby et al: II Task• 3 of the 4 dimensions determine categories• Not readily verbalisable

Cat A

Cat B

IMPASS & CL: Study 5 cont.• Ss did 2 sessions one week apart• 2 RB tasks in one session and 1 II task

in the other• Also measured paired-associate

episodic memory as before and working memory performance in each subject

• Variety of personality measures

Working Memory (WM) Task

• Memory set scanning task• A set of 6 letters presented

simultaneously for 2.5 secs• Y/N testing with 12 letters• 10 sets used each quasi-randomly

selected from 24 letters (no O or L)• Overall % correct recorded

Preliminary Results 1Regressions predicting CL %correct

IMPASS

SCHIZO

RB1 0.06 -0.11

RB2 0.06 0.11

II 0.17† -0.20*

† p=0.06

Preliminary Results 2Regressions predicting CL %correct

WM %corr

PA #corr

RB1 0.07 0.14

RB2 0.04 0.19*

II 0.37** -0.05

Preliminary Results 3Regressions predicting %correct on II task:

IMPASS

SCHIZO

WM

%corr

PA

#corr0.13 -0.20* 0.33** 0.02

Interpretation & Conclusions: 4• Notionally II task strongly dependent

on WM (more so than notionally RB task)

• Positive correlation between IMPASS and CL performance replicated; reflects contribution of WM on task? (for rules or exemplars?)

• Negative relationship between positive schizotypy and CL performance also replicated and independent of WM

IMPASS and CL: Study 6• Study by Ian Tharp using an

information integration task from the Ashby/Maddox stable

• Stimuli were lines which varied in length and orientation

• 100 training exemplars with feedback, each presented twice

• 48 healthy male subjects (not all students)

• 4 different IMPASS measures

II Task Structure

0 100 200 300 400 5000

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

Length in Pixels

Orienta

tion in R

adia

ns

Correlations with 4 IMPASS measures

IMPNON EPQ-P SSS BAS-FS

% Corr

-0.32* -0.17 -0.30* -0.27

Contrasts with previous 4 studies where correlations were positive

Preliminary Modelling 1• Fit a “General Linear Classifier” model

(i.e., discriminant function D) to responses of each individual subject

• D = b1*length + b2*orient + c0

• Relative values of b1 and b2 are informative w.r.t. task strategyunidimensional b1>>b2 or b2>>b1

bidimensional b1 b2

Preliminary Modelling 2Converted b1 and b2 into strategy index

Strategy Index

1.00

.94

.88

.81

.75

.69

.63

.56

.50

.44

.38

.31

.25

.19

.13

.06

0.00

Fre

qu

en

cy

10

8

6

4

2

0

UNIBI

Correlations with Strategy Index

% Corr IMP-NON

EPQ-P SSS BAS-FS

Strat. Index

-0.51** 0.26 0.31* 0.23 0.16

A unidimensional strategy harms performance and is favoured by high IMPASS subjects

1. Positive effect of involvement of working/episodic memory

2. Effect of preference for simple unidimensional rules (can be positive or negative)

Interpretation & Conclusions: 5

Perhaps CL performance of high IMPASS Ss reflects two distinct processes:-

IMPASS and CL: Study 7• Study with Luke Smillie• Within-Ss design using 2 CL tasks

Information integration: Occupational Selection taskEpisodic memory task:Good vs. Bad Numbers task

• 102 Australian psychology students• 2 different IMPASS measures

Numbers Task• Quasi-randomly 6 2-digit numbers

designated “good” and 6 2-digit numbers are designated “bad”

• Number selection avoids obvious rules• Go vs. no-go responses with feedback• Explicitly instructed to memorise• 96 trials• Predict positive correlation with IMPASS

Occupational Selection Task: OST

• S presented with “ratings profiles” of candidates on 5 job attributes

• S has to decide whether to hire• 100 trials with feedback• 50 suitable candidates who should be

hired and 50 unsuitable• Instructed: “use only the ratings

profiles, each attribute reliably but modestly related to suitability”

OST: Stimuli and Predictions• On each dimension, ratings were

normally distributedSuitable mean = 55 s.d. = 18Unsuitable mean = 40 s.d. = 18

• Correlations between …dimension and category 0.37-0.43dimensions 0.2-0.6

• Predict negative correlation with IMPASS

OST: Preliminary Results

• Correlations between task performance (d’) and IMPASS

EPQ-P EPP-SS

Numbersd’

0.16* 0.16*

OSTd’

-0.14* -0.14*

General Conclusions• Modest but reliable associations

between personality traits and CL performance

• These relationships depend on type of CL task used in a relatively predictable way

• Findings contribute to the multiple systems view of CL performance