Probabilistic categorization, selective attention, and sensitivity to correlated cues

1
Probabilistic categorization, selective
attention, and sensitivity to correlated cues

• Daniel Little
• Stephan Lewandowsky
• University of Western Australia

2
Selective Attention

• In category learning, it is implicit that people
  attend to valid cues
• Tacit assumption of GCM (Nosofsky, 1986)
• Asymptotic behavior of a connectionist network
  (e.g., ALCOVE Kruschke, 1988)

3
Selective Attention

• Distribution of attention can explain many
  benchmark categorization phenomena
• Blocking
• Highlighting
• Relative learning difficulty
• Shift learning
• However, several extensions to basic selective
  attention mechanisms have been proposed

Attention
4
Selective Attention

• Rapid shifts of attention in response to error
  (RASHNL, EXIT)
• When an error is made, attention is rapidly
  shifted to other stimulus dimensions
• Learning effects (e.g., learned inattention
  Kruschke, 2001)
• Interactions between cue salience and cue
  validity (Kruschke Johansen, 1999)
• Inverse base rate effect (Kruschke, 2005)

5
Probabilistic Categorization

• Probabilistic assignment of items to categories
• Items will appear in both categories
• Perfect performance is impossible

A
CORRECT
6
Probabilistic Categorization

• Probabilistic assignment of items to categories
• Items will appear in both categories
• Perfect performance is impossible

A
WRONG
7
Probabilistic Categorization
Probabilistic Feedback P(A) .75
75
25
8
Probabilistic Categorization

• Maximizing always respond to the category with
  the
• higher payoff

75
25

• Probability Matching allocate responses in
  proportion
• to the payoff
• People in probabilistic tasks typically
  probability match (e.g., Shanks et al.,
• 2002) they do so in the present probabilistic
  task

9
Probabilistic Categorization

• Probabilistic assignment of items to categories
• Items will appear in both categories
• Perfect performance is impossible
• In probabilistic categorization the level of
  unavoidable error is increased
• Even an optimal distribution of attention can not
  eliminate error completely

10
Probabilistic categorization

• Increased error should result in increased
  attention shifting
• Evidence for this comes from irrational
  cue-competition effects
• Irrelevant features might be perceived as being
  valid
• Relevant features might be perceived as not being
  valid (Castellan, 1973 Edgell et al. 1992).
• Will attention shifting increase sensitivity to
  non-relevant correlated cues?

11
Correlated Cues

• What are correlated cues?
• Birds ? have wings can fly have feathers
  have beaks lay eggs, etc.
• Small birds tend to sing, large birds do not

12
Correlated cues

• Why are correlated cues important?
• All of these correlations exist in the
  environment
• To not take advantage of them would be irrational
  (wouldnt it?)
• The results about whether people have knowledge
  about correlated cues is mixed

13
Correlated cues

• People are sensitive to correlated cues in some
  tasks
• Incidental learning (i.e., rate this person)
• Wattenmaker (1991, 1993) Wattenmaker et al.
  (1995)
• Feature inference
• Chin-Parker Ross (2002)
• But not in intentional category learning
• Wattenmaker (1993), Chin-Parker Ross (2002)

14
Experiment

• If probabilistic assignment of stimuli to
  categories increases attention shifting
• this increased attention shifting might lead
  people to notice an otherwise irrelevant
  correlation

Probabilistic Feedback
Correlational Sensitivity
?
Attention
?
15

• Category Learning Experiment
• Two categories
• Two conditions
• Deterministic
• Probabilistic

Correct
Wrong
16
X
Y
Z
Color
17
Current Experiment

• Probabilistic Condition

Category A P(AStimulusA) .75
18
Current Experiment

• Probabilistic Condition

Category A P(AStimulusB) .25
19
Current Experiment
Type II Shepard et al. (1961) Exclusive-or

• Deterministic Condition

Category A P(AStimulusA) 1.0
20
Current Experiment

• Z Color
• Perfectly correlated during training
• Are people sensitive to non-relevant correlated
  cues?

21
Current Experiments
CS 0 No switching when color changes
NEW

• OLD

INDEX OF CORRELATIONAL SENSITIVITY The total
proportion of times that a participant switches
their response between an old item and a new
item (given that all that changes is the status
of the correlation)
CS 1 Switching every time color changes
A
B
A
22
Experiment

• Are people sensitive to correlated cues in
  categorization?
• Yes, but only in the probabilistic feedback
  condition
• Replicated previous non-sensitivity in the
  deterministic condition

23
N 10
N 26
24
Attentional Diffusion

• Are people sensitive to correlated cues in
  categorization?
• Yes, but only in the probabilistic feedback
  condition
• So youre more likely to learn that small birds
  sing if on occasion you see a small creature that
  sings which is not a bird

25
(No Transcript)
26
Relationship
Probabilistic Feedback
Correlational Sensitivity
Diffuse Attention Profile
27
Modeling

• To explore this, we fit three exemplar models
  that differ in how attention is allocated
• GCM (Nosofsky, 1986)
• Basic Exemplar Representation
• ALCOVE (Kruschke, 1992)
• Exemplar Representation connectionist learning
• Attention is learned across trials
• RASHNL (Kruschke Johansen, 1999)
• Extension of ALCOVE, designed explicitly for
  probabilistic tasks
• Salience parameter
• Acts as an additional multiplier on psychological
  distance

28
Modeling

• To explore this, we fit three exemplar models
  that differ in how attention is allocated
• GCM (Nosofsky, 1986)
• Basic Exemplar Representation
• ALCOVE (Kruschke, 1992)
• Exemplar Representation connectionist learning
• Attention is learned across trials
• RASHNL (Kruschke Johansen, 1999)
• Extension of ALCOVE, designed explicitly for
  probabilistic tasks
• Salience parameter
• Acts as an additional multiplier on psychological
  distance

29
Modeling

• Parameters optimized to each individuals final
  training performance
• Transfer predictions generated from each model
• Only the GCM was able to predict
• Transfer utilization
• Correlational sensitivity

30
GCM - Predicted Utilization
Deterministic Condition
Probabilistic Condition
Predictions generated after various points during
training Learning simulated in the GCM by
adding a new exemplar after every trial
YZ YC
YZ YC
31
Correlational Sensitivity
Data
32
Correlational Sensitivity
Data
33
Modeling

• Only the GCM predicts both the deterministic and
  probabilistic condition
• Used the GCM to explore the relationship between
  probabilistic feedback, attention shifting and
  sensitivity to correlation

34
Exploration

• Fit probabilistic training with deterministic
  data
• Assumes maximizing rather than probability
  matching
• Attention profile more focused when we assume
  maximizing and not probability matching

35
(No Transcript)
36
Relationship
Probability Matching
Probabilistic Feedback
Correlational Sensitivity
Diffuse Attention Profile
37
Exploration

• If probability matching is associated with a more
  diffuse attention profile
• and given that the probabilistic condition
  showed increased correlational sensitivity but
  the deterministic condition did not
• then we might expect to find an empirical
  relationship between probability matching and
  correlational sensitivity

38
Exploration

• probability matching and correlational
  sensitivity?
• PM Score 1 ? Maximizing
• PM Score .75 ? Probability Matching
• PM Score .50 ?
• Chance

39
Relationship
Probability Matching
Probabilistic Feedback
?
Correlational Sensitivity
Diffuse Attention Profile
40
Exploration

• What happens as responding moves from
  deterministic to probabilistic?
• Create a continuum of parameters from
  deterministic to probabilistic by linearly
  interpolation
• All parameters
• Attention parameters only
• The probability matching behavior of the model is
  constant, but attention is made more diffuse

41
Exploration
42
Relationship
Probability Matching
Probabilistic Feedback
Correlational Sensitivity
Diffuse Attention Profile
43
Thank you

• Probabilistic feedback leads to
• Probability matching
• Diffuse attention profile
• Diffuse attention profile leads to correlational
  sensitivity
• So youre more likely to learn that small birds
  sing if on occasion you see a small creature that
  sings which is not a bird
• because you have to attend to more than size and
  song.

44
For more information

• Little, D. R. Lewandowsky, S. (in press).
  Beyond non-utilization Irrelevant cues can gate
  learning in probabilistic categorization. Journal
  of Experimental Psychology Human Perception and
  Performance.
• Little, D. R. Lewandowsky, S. (2008).
  Probabilistic Feedback Increases Sensitivity to
  Correlated Cues. Manuscript submitted for
  publication. University of Western Australia.
• Lewandowsky, S., Little, D. R., Kalish, M. L.
  (2007). Knowledge and expertise. In F. T. Durso,
  R. Nickerson, S. Dumais, S. Lewandowsky, T.
  Perfect (Eds.). Handbook of applied cognition,
  2nd Ed. (pp. 83 110). Chicester Wiley.