Modelling Language Evolution Lecture 3: Evolving Syntax

1
Modelling Language EvolutionLecture 3 Evolving
Syntax

• Simon Kirby
• University of Edinburgh
• Language Evolution Computation Research Unit

2
Evolving the ability to learn syntax (Batali 1994)

• A standard recurrent network does not seem to
  be able to learn syntax without some help
• Elman provides this help via incremental memory
• The network comes pre-setup to help it learn
  syntax
• i.e., our model of an individual is born with a
  working memory that grows over time
• Does this correspond to an innate
  prespecification for language learning?

3
Where do innate abilities come from?

• If an organism has some innate predisposition
• and that predisposition is functional, how do
  we explain it?
• Darwinian natural selection seems appropriate.
• Could we model natural selection?
• Can we evolve a syntax learner? (as opposed to
  building one by hand?)

4
What things about a network could be innate?

• Many features of networks could be thought of as
  innately determined
• The length of time before context units are
  blanked
• The shape of the activation function
• The number of nodes in the hidden layer
• Batali suggests the initial connection weights.
• Normally, these are random but what if they
  were specified by genes?

5
How to model an organism
0.2 -1.3 0.05 0.9 -0.5 0.001 0.1
GENOTYPE
development
PHENOTYPE

• The model has genes, which are expressed as a
  phenotype.
• The phenotype is simply the initial state of a
  network (before learning).

6
How to evolve organisms

• Crucial aspects of evolution
• A population of organisms (with varying
  phenotypes)
• A task which they are trying to succeed at
• A measure of how fit they are at this task
• A way of selecting the fittest
• A way of allowing the genes of the fittest to
  survive
• A mechanism for introducing variation into the
  gene pool
• Various techniques to model all of this (i.e.,
  Genetic Algorithms, Artificial Life etc.)

7
Batalis model of evolution

• Each organism (or agent) has its weights set by
  genes
• The agents then trained on some language
• The agents error is used to assign fitness
• Only the top third of the population is kept
• The top third have their weights reset to what
  their genes specify
• Each agent gives birth to two new agents with
  approximately the same genes (i.e., genes are
  mutated)
• Go to step 1.

8
The language task

• One of the simplest languages that involves
  embedding is
• ab
• aabb
• aaaaaaaabbbbbbbb
• aaaaaaaaabbbbbbbb
• What machinery would you need to recognise
  strings from this language?
• Minimally a simple counter
• Can an SRN with random initial weights learn this
  language?

9
Performance of a trained (but non-evolved) net

• Networks fail to learn to count (although some
  aspects of the language are learnt).

10
Evolving a better network

• Batali used a population of 24 nets (initially
  with genes specifying random weights)
• Evolved using a fitness function based on
  abilityat after training
• After 150 generations, the networks were better
  at learning the task
• They evolved initial weight settings that made
  learning syntax possible

11
Evolved network performance
a
b
sp
rec
12
(No Transcript)
13
Issues that remain

• What is learning doing?
• If language is always the same, the networks
  could eventually end up with the whole thing
  innate (and not need learning at all!)
• What would happen if the networks were trained on
  a class of languages?
• Initial weights are a different type of
  innateness than Elmans. Can Batali also explain
  the critical period?

14
Is evolution just the same as learning?

• We can think of a fitness landscape just like an
  error surface
• What are the differences?
• Does evolution do gradient descent?