Interindividual Differences in Discrimination Learning

1
Inter-individual Differences in Discrimination
Learning

• Verena Schmittmann Maartje Raijmakers
• University of Amsterdam

2
Different strategies in Rule/Feedback Learning

• Focus shifts from positive to negative feedback
  possibly different strategies (e.g., van
  Duijvenvoorde, et al 2008)
• Rule learning in adults (e.g., Sirois et al,
  2006 Mutter et al, 2006)
• Discrimination learning task Evidence for
  different modes of learning (e.g., Kendler, 1979
  Raijmakers, Dolan and Molenaar, 2001)

3
Simple discrimination learning task

• Stimuli differ on two binary valued dimensions
• Form (Circle/Triangle)
• Size (Small/Large)
• Feedback provides information on correct
  feature Circle, Triangle, Small, or Large

4
Underlying learning processes - sequence of
responses of one child (red)
correct
error
correct
underlying sudden learning process
error
correct
underlying gradual learning process
error
5
Data

• two studies ( N230, 4-20 years and N300,
  4-14 years)
• from each participant, sequence of accuracies on
  all trials was recorded, e.g.,011000110101111111
  11 0110111111111

6
Fast Hypothesis Testing

• Fast learning, well described by model of
  hypothesis testing

1
lp learning parameterblue transition
probabilitiesred probability of correct response
lp
Adapted from original CI model, Bouwer
Trabasso, 1964
7
Development
Estimated Proportion of Fast Hypothesis
Testers (bars 95CI)
Age Group
8
Slow Learning

• One-trial learning from chance level to near
  perfect performance
• Less efficient than choosing hypothesis at random
  from pool of features upon every error
• Likely inefficient hypothesis testing

Trial of Learning
ErrorProb.
Blocks of Four Trials
9
Why trial-by-trial modeling?

• theory-based
• averaging may distort results, change inferences

Probability correct
10
Discrimination shift learning

• after learning criterion, reversal shift of
  reinforcement contingencies
• different learning strategies in shift phase?
• same strategy in both phases of task? expected
  rational learning in initial phase -gt rational
  learning in shift phase

11
Discrimination Shift Model
12
Results Shift Learning

• Two modes of reversal shift learning
• Fast initial learning may be paired with slow
  shift learning (neg. transfer especially in the
  youngest age group), and with fast shift learning
• Slow initial learning may be paired with fast
  shift learning (pos transfer especially in the
  older age groups)
• Response accuracy in applying the reversed rule
  increased with age, and is below criterion
  accuracy in the youngest age groups

13
Number of trials to fixed learning
criterionInitial learning (x) by shift learning
(y)in 4 age groups
traditional measure overestimates trial of
learning in young children
14
Questions

• Do some 4-5 year olds employ a fast hypothesis
  testing strategy, or do they have a preference
  for the correct feature?
• Relation of learning strategy to - dimension
  preference? - attentional control? - working
  memory?

15
Study II

• N302 4 to 14 years, 5 age groups
• Discrimination learning task, relevant dimension
  non-preferred dimension
• Attentional control/ Inhibition Eriksen flanker
  task (e.g., Ridderinkhof et al, 1995)
• Working memory spatial working memory task (v.
  Leijenhorst et al. 2007)
• Conditional reasoning task

16
Mixture model
Initial ProbabilitiesL .053E .5C
.447
Initial ProbabilitiesL 0P 1
Initial ProbabilityG 1
Markovian Learning models fitted to
trial-by-trial accuracy data by ML estimation
using the R-package depmix (Ingmar Visser)
17
DevelopmentProportions of learning strategies
by age group
F Fast Hypothesis Testing S Slow Learning N
Non-learning
Estimated Proportion of Strategy (bars SE)
Age Group
18
Preference Learning Strategy 4-5 year-olds
Estimated Proportions of Strategies
Strategy
proportion did not differ sig.from 0 gt fixed
19
Hierarchical Regression AnalysisAge, Attentional
Control Working Memory

• Age significant predictor of all learning
  strategies
• Working Memory and Attentional Control measures
  significant predictors of fast hypothesis testing
  and of non-learning

20
Conclusions

• statistical modeling support for existence of
  mixture of learning processes
• no evidence for incremental discrimination
  learning in young children
• Evidence for fast hypothesis testing in 4-5
  year-olds without strong preference
• preference for irrelevant dimension hurt learning
  in youngest children (4-5 yrs)
• working memory and attentional control predicted
  learning strategy
• trial-by-trial modeling of learning processes can
  reveal relevant information

21
Acknowledgments

• Maartje Raijmakers
• Han van der Maas
• Ingmar Visser
• Raoul Grasman
• Linda van Leijenhorst
• Mariette Huizinga
• undergraduate students
• participants