Learning

1
Learning Memory

• 3. Human contingency learning

2
Human contingency learning

• Goal Understanding human learning
• Key to understanding of the human mind
• Contingency learning is starting simple
• Commercial/social implications
• Traders predictions of stock market
• Stereotypical judgements about minorities
• The limits of evolutionary theory

3
Human contingency learning
Defining contingency
Accuracy of judgements

• Essential
• Wills, Ch. 2
• Shanks, Ch. 2
• Suppl
• Wills, Ch. 7

The effect of experience
Contingency chi-square 1117-5
Trial-order effects
Predictive redundancy
Rescorla-Wagner model
Before
After
During
4
Contingency chi-square
5
Contingency Thunder lightning
Outcome
Thunder
No thunder
6
1
Lightning
Cue
5
1
No lightning
6
Contingency Delta P
O
O
P(OC) 6 / (61) 0.86
C
C
P(OC) 1 / (51) 0.17
?P 0.86 0.17 0.69
?P P(OC) - P(OC)
7
Negative Delta P

• ?P P(OC) P(O-C) 0 1 -1
• ?P is maximally negative.
• Cue is a perfect predictor of the absence of the
  outcome.

8
Zero Delta P

• ?P P(OC) P(O-C) 0.5 0.5 0
• ?P is zero.
• Cue and outcome are non-contingent.
• Common cues are not necessarily predictive cues.

9
Subjective contingency

• Subjective ratings and ?P
• ?P is an objective degree of relatedness (in
  simple situations).
• Significant evolutionary pressure for accurate
  contingency estimation.
• Subjective judgements should conform to ?P.
• Is this actually the case?

10
Accuracy of contingency judgements
11
Accuracy of contingency judgements

• Wasserman, Elek, Chatlosh Baker (93)
• One action - Telegraph key
• One outcome - Light flash
• Free operant procedure
• 4 minutes per condition
• 25 conditions
• P(OA) 0, 0.25, 0.5, 0.75, 1
• P(O-A) 0, 0.25, 0.5, 0.75, 1
• Judgement on scale of -100 to 100

12
Wasserman et al.
13
Lopez Shanks (Shanks 1995)

• Tanks colour-sensitive mines.
• One cue - Tank is coloured.
• One outcome - Tank explodes.
• 40 trials.
• Judgements (-100 to 100) every 5 trials.
• Four conditions
• 0.75/0.25, 0.25/0.75, 0.25/0.25, 0.75/0.75

14
Lopez Shanks
15
Why should experience matter?
?P P(OC) P(O-C) 0.5 0.5
0
?P P(OC) P(O-C) 0.5 0.5
0

• Perhaps theres more to contingency judgements
  than delta P?

16
Trial order

• Chapman (1991)
• Simulated medical diagnosis
• Four cues - Symptoms A,B,C and D.
• One outcome - Disease.
• Three discrete stages of presentation

17
Trial order

• B rated as more negative than D.

18
Learning and surprise

• Baseline cues not associated with disease.
• A ? O Surprising - no disease expected
• AB ? no O Surprising - disease expected
• CD ?no O Unsurprising - no disease expected
• C ? O Surprising - no disease expected.
• Learning is driven by surprise.
• More is learned about B than about D.
• Hence, ratings are different.

19
Predictive redundancy

• Shanks (1991)
• Medical diagnosis 6 symptoms, 2 diseases

20
Predictive redundancy

• For any given outcome, we are more likely to
  remember the most reliable predictors than the
  least reliable.
• Explained by the Rescorla-Wagner theory (see
  lecture notes, readings, and next weeks
  tutorial).

21
Summary

• Whilst ?P is an objective measure of contingency,
  our judgements only correspond to it in simple
  situations of which we have plenty of experience.
• Where experience is slight, or heterogeneous over
  time, or where an outcome has more than one
  possible cause, ?P and our judgements differ.
• The difference stems from the fact that
  contingent relationships are learned
  incrementally, whilst ?P is an after the event
  reflection on the contents of a perfect memory.

22
What next?
Classical vs. FR 1121-2

• Essential
• Wills, Ch. 2
• Shanks, Ch. 2
• Suppl
• Wills, Ch. 7

Proto. theory evid. 1121-2
Instance theory evid. 1121-2
McC R 85 1121-2 (VRR)
Semantic Nets 1121-2 2021-7