Inductive learning

1
Inductive learning

• Concept acquisition
• Causal inference
• Language acquisition

2
Induction vs. deduction

• Induction
• Deciding what is to be learned
• Figuring out new computer applications
• Inference and generalization
• Deduction
• Syllogistic reasoning
• Conclusions are necessary if the premises are
  true.
• Logical errors are possible.

3
Deductive reasoning

• All dogs have tails.
• Bo is a dog.
• Therefore, Bo has a tail.
• Muffy has a tail.
• Therefore, Muffy is a dog.
• Koko has no tail.
• Therefore, Koko is not a dog.

4
I. Concept acquisition

• Hypothesis testing
• Natural concepts or categories
• Schema formation
• Prototypes or exemplars

5
Concepts

• Rational categories
• truck, bicycle, hovercraft
• tomorrow, Feb. 30, Marsday
• 25, 81, 121
• Are they learned by association (Hull, 1920)?
• Chinese radical categories
• Learning without awareness

6
Bruner, Goodnow Austin (1956)

• Find blix

7
Hypothesis testing

• What is the rule?
• What rule fits prior experience?
• Test new instances against that rule.
• Evaluate hypothesis.
• Introduces discontinuous learning.
• All-or-none learning and the backwards learning
  curve
• Group data vs. individual data

8
Natural categories

• Vague boundaries
• Degrees of membership Tomatoes
• Disagreement about membership
• McCloskey and Glucksberg (1978) stroke, leech,
  pumpkin
• Typical members have more of the features of a
  category robin is a bird, penguin is sort of a
  bird.

9
Schemas and exemplars

• Schemas Specific rules, prototypes. Very
  rational. (Gluck Bower, 1988)
• Exemplars Group assignment is based on
  similarity to other examples, not on a defined
  category. Very empirical. (Medin Schaffer,
  1978)
• Instruction, hypotheses, schemas, or exemplars?
  It depends on the situation.

10
Concept acquisition

• Direct instruction, as in education
• Hypothesis testing where there are rigid rules to
  be discovered
• Schema theory for categories with vague
  boundaries
• Exemplar theory for categories with scattered
  members

11
II. Causal Inference (Einhorn Hogarth, 1986)

• Statistical cues
• Spatial and temporal contiguity cues
• Kinematic cues
• Complex systems

12
Statistical cues

• Contingency analysis The 2 x 2 contingency
  table, probability analysis

Effect
Humans are most sensitive to a, next to b and c,
and least to d in conscious hypothesis testing.
-
13
Spatial and temporal contiguity cues

• Lightning and thunder
• Spilled milk and kidsbut broken vase and cats?
• Bullock, Gelman Baillargeon (1982)
• If equidistant, cause closer in time was chosen
  by 65 of the people
• If equitemporal, cause closer in space was
  chosen by 100 of the people
• If one is closer in time, the other closer in
  space, 70 chose the cause closer in space.

14
Prior causal models

• Wave theory Inverse square law
• Prediction of linear acceleration Depression
  effect

15
Kinematic cues

• Billiard-ball movement
• timing
• velocity
• angle of reaction
• Naïve physics models
• Improve if education and experience are available
• May be parabolic The c-shaped tube problem. Why?

16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
Complex systems

• Generalization from previous experience, eg.
  Computer programs
• Meaning of stimuli
• Control metaphors Enterprise

21
III. Language acquisition

• Vocabulary by association?
• Grammar by induction, less by rules
• Holophrastic speech
• Past-tense acquisition
• Deaf children learn language, even if it is not
  taught.

22
Language A developmental process

• LAD? (Chomsky)
• Critical periods
• Songbirds
• Foreign language learning
• Aphasia and recovery
• Language universals

23
Animal language learning?

• Washoes 132 signs, some word order variation for
  subject-object distinction
• Premacks Sarah Plastic shapes
• Terraces critique Nim Chimpsky
• Kanzi, the bonobo Observational learning