Iterated Learning in birds, humans and machines

1
Iterated Learning in birds, humans and machines

• Evolution of Song Culture in the Zebra Finch
  (Feher et al'08, Evolang)
• Innateness and culture in the evolution of
  language (Kirby et al'07, PNAS)

2
Outline

• What is iterated learning?
• Why is it relevant for human language?
• Appearance of design (Pinker Bloom'90)
• Innateness vs. constraints on variation
• Bayesian ILM sampling vs. maximizing (Kirby et
  al. 2007)
• Why is it relevant for bird song?
• Feher et al'08 Evolution of song culture
• Future experiments?

3
Iterated Learning
4
Iterated Learning
5
Iterated Learning

• Learners learn from learners
• Cultural transmission
• Claim in cultural transmitted systems one has to
  be extra careful to classify things as nature or
  nurture, or as adaptation or side-effect

6
Relevance for Linguistics

• Which mechanisms underlying language are unique
  for humans and unique for language?

7
Poverty of the stimulus

• (1) The man is mean
• (2) Is the man mean?
• (3) The man who is feeding a donkey is mean
• (4) Is the man who is feeding a donkey _ mean?
• (5) Is the man who _ feeding a donkey is mean?

8
1980s nativism

• Any aspect of language that the speaker knows
  must either be learnable from positive evidence,
  that is to say, through exposure to sentences of
  the language, or be part of the innate equipment
  of the human mind (Cook, 1983)

9
1980s nativism

• Any aspect of language that the speaker knows
  must either be learnable from positive evidence,
  that is to say, through exposure to sentences of
  the language, or be part of the innate equipment
  of the human mind (Cook, 1983)
• Language shows signs of complex design for the
  communication of propositional structures, and
  the only explanation for the origin of organs
  with complex design is the process of natural
  selection. (Pinker Bloom, 1990)

10
Universal Grammar
Language-specific adaptations
Language acquisition device Initial state
Constraints on variation
11
Universal Grammar
Language-specific adaptations
Language acquisition device Initial state
Constraints on variation
Iterated Learning appearance of design can
occur even without any biological
evolution (Kirby, 1994/2000)
12
Example

• Is speech perception uniquely human and uniquely
  linguistic?
• (e.g. Pinker Jackendoff, 2005, Cognition reply
  by Fitch, Hauser, Chomsky, 2005)
• Speech is Special (Liberman, 1950s)
• - Categorical Perception...... not unique
  (Kuhl Miller'75)
• - Perceptual Magnet Effect..... not unique
  (Kluender et al'98)
• Speech is phenomenological different
• Speech neurologically doubly dissociates
• Speech is preferred by neonates
• Speech perception in humans differs from other
  primates in micro-features (exact vowel
  boundaries, attention to voice on-set time etc.)

13
Why are humans so good at making linguistically
relevant auditory distinctions? (i) auditory
perception has evolved, to allow humans to better
perceive speech sounds (ii) auditory perception
has remained unchanged (modulo drift and
side-effects) appearance of design is solely due
to cultural evolution (iii) auditory perception
has remained unchanged (modulo drift and
side-effects) appearance of design is due to
genetic evolution of articulation, learning,
UG (iv) auditory perception has evolved, such
that the outcome of cultural evolution is
biologically adaptive.
14
4-matrix model sending
(Hurford, 1989)
S
15
4-matrix model receiving
(Hurford, 1989)
R
16
4-matrix model confusion
(Nowak Krakauer, 1999)
U
17
4-matrix model similarity
(Zuidema Westermann, 2003)
V
18
4-matrix model payoffs
(Zuidema Westermann, 2003)
19
(No Transcript)
20
(No Transcript)
21
S
U
R
V
22
4-matrix model payoffs
(Zuidema Westermann, 2003)
23
4-matrix model payoffs
(Zuidema Westermann, 2003)
Hearer behavior (learned)
Speaker behavior (learned)
Values of alternate interpretations (assumed
constant)
24
4-matrix model payoffs
(Zuidema Westermann, 2003)
Confusability of alternate signals articulation,
acoustics, perception (assumed constant)
25
Reliability of recognition
Acoustic feature
(We choose random values from 0,1 on the
diagonal of U).
26
Agents in a population maximize payoff by
adapting S and R (local hillclimbing) much like
the naming game (Steels'95)
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
S
R
27
(No Transcript)
28
U
S
T
R
29
U
S
T
R
30
U
S
T
R
31
Negative conclusions

• There is a tight fit between the perceptual
  characteristics and the speech code
• But in this model, the perceptual characteristics
  did not evolve/adapt
• Rather, through cultural evolution, the language
  has evolved to match perception
• The model thus shows that appearance of design
  cannot be taken as evidence for (genetic)
  adaptation.

32
Formal models

• Mathematical models of Gold (1967) and Nowak et
  al (2001, Science) have been interpreted as
  proofs of an extensive, innate UG
• Iterated learning perspective also reveals
  problems with these interpretations (Zuidema,
  2003, NIPS)

33
Universal Grammar
Language-specific adaptations
Language acquisition device Initial state
Constraints on variation
34
Binary learnability
Language-specific adaptations
Language acquisition device Initial state
Constraints on variation
Chomsky Not all language that the UG allows need
to be learnable, but unlearnable languages will
not occur in the next generation.
35
Graded learnability
?
Language-specific adaptations
Language acquisition device Initial state
Constraints on variation
Kirby et al (03), Zuidema (03) graded
learnability implies very nontrivial
relation But Griffiths Khalish (2006)
stationary prior distribution
36
Graded learnability
37
Stationary distribution
P(GD)P(DG) P(G)