Filetype PDF | Posted on 24 Sep 2022 | 3 years ago
Authentication
digital resources
130x Filetype PDF File size 0.54 MB Source: cocosci.princeton.edu
Available online at www.sciencedirect.com
ScienceDirect
Iterated learning and the evolution of language
1 2 1
Simon Kirby , Tom Griffiths and Kenny Smith
Iterated learning describes the process whereby an individual Iterated learning: The process by which a behaviour
learns their behaviour by exposure to another individual’s arises in one individual through
behaviour, who themselves learnt it in the same way. It can be induction on the basis of observations
seen as a key mechanism of cultural evolution. We review behaviour in another individual
of
various methods for understanding how behaviour is shaped by who acquired that behaviour in the same
the iterated learning process: computational agent-based way.
simulations; mathematical modelling; and laboratory
experiments in humans and non-human animals. We show how For example, we induce the particular properties of our
this framework has been used to explain the origins of structure language by being exposed to the linguistic behaviour of
in language, and argue that cultural evolution must be other individuals in our speech community. Our resulting
considered alongside biological evolution in explanations of language in turn leads to linguistic behaviour that shapes
language origins. the language of further individuals, leading to the possib-
ility of cultural evolution by a process of repeated induc-
tion and production of behaviour. In this paper we survey
Addresses simulations, mathematical models and experiments all
1University of Edinburgh, Edinburgh, Scotland, United Kingdom
2University of California at Berkeley, USA pointing towards the same underlying hypothesis: that
key structural design features of language have their
the
Corresponding author: Kirby, Simon (simon@ling.ed.ac.uk) explanation in the fact that language is culturally trans-
mitted in this way [4 ,5–7]. The rarity of this kind of
design in natural communication may appear to be
explained as a consequence of the rarity of iterated
Current Opinion in Neurobiology 2014, 28:108–114 learning. However, as we will argue at the end of this
This review comes from a themed issue on Communication and review, the vocal productions of some other species —
language
most notably, songbirds [8 ] — also evolves culturally via
Edited
by Michael Brainard and Tecumseh Fitch iterated learning. This opens up an intriguing avenue for
study, and also raises important questions
comparative
about the differences in the design features of song and
language.
http://dx.doi.org/10.1016/j.conb.2014.07.014 Agent-based simulation
0959-4388/# 2014 Published by Elsevier Ltd. Foundational work by Hurford [9] sparked interest in
computational simulation as a tool for modelling the
biological and cultural evolution of language. Following
Hurford’s lead, the earliest work in this area sought to
explain the role of interaction and negotiation [10 , 11] or
biases of learners [12,13] in shaping communication sys-
Introduction: can culture explain structure? focusing in particular on the conditions under which
tems,
Language exhibits striking structural design features that communicatively optimal, socially learnt communication
mark it out as extremely unusual among communication systems would emerge. Subsequent efforts were directed
systems in nature. In particular, utterances in a language towards an explanation of how linguistic structure can
are constructed out of sub-parts — phonemes, mor- arise as a consequence of iterated learning. While inter-
phemes, words, phrases — that are reused and recom- action and learning bias play a role in this process [14,15],
bined in systematic ways. Because of the apparent much of this work emphasises the role of the learning
uniqueness of this design, and because it enables the bottleneck [4,15–19] in driving the evolution of structure:
open-ended expressive potential of human language, language learners must attempt to learn a large or
structure has been a primary target for expla- expressive linguistic system on the basis of a
linguistic infinitely
nation by evolutionary linguists and cognitive science relatively small set of linguistic data. A key finding is that
more generally [1–3]. compositional languages (in which the meaning of a com-
plex expression is composed of the meanings of parts of
In addition to exhibiting structure, language is one of a that expression) emerge from holistic (i.e. unstructured)
rare set of behaviours that persists through a particular languages as a result of repeated transmission through the
kind of cultural transmission: iterated learning [4]. learning bottleneck: language structure appears as an
Current Opinion in Neurobiology 2014, 28:108–114 www.sciencedirect.com
Iterated learning and the evolution of language Kirby, Griffiths and Smith 109
adaptive response by language itself to the problem of being depends on how many people understand that language,
transmitted through a narrow bottleneck, since the and children can end up speaking different languages
presence of compositional rules enables a learner to infer from their parents. This framework can be used to rig-
from a small sample rules underpinning the entire orously answer questions about, for instance, how con-
language. strained language learning needs to be in order to
guarantee that a population will end up speaking the
Another rich seam of modelling work looks at the emer- same language, and to what extent this coherence threshold
gence of systematicity in phonological systems through can drive the evolution of an ever-more restrictive
communicative interaction and iterated learning. For language faculty [29].
example, De Boer [20,21] looks at the cultural evolution
of vowel systems, demonstrating that universal features of Perhaps as a consequence of their origins in biological
the organisation of vowels in the world’s languages can evolution, these models made very weak assumptions
arise through repeated interaction between simulated about the transmission process itself: no language is
agents under certain reasonable articulatory and percep- easier or harder for learners to acquire than any other.
tual constraints. Models by Oudeyer [22,23], Wedel As a consequence, the driving force in the dynamics was
[24,25], and Zuidema and De Boer [26], despite very the effect of fitness — of being able to communicate
different underlying assumptions about the cognitive effectively with others — rather than learning. To
machinery involved, show that the process of repeated explore the effects of transmission more directly,
learning and use of a sound system can lead to the Griffiths and Kalish [33 ] developed the first mathemat-
emergence of systematic organisation of sequences of ical characterisation of the results of iterated learning,
sounds, as well as the organisation of those sounds them- based on analysing vertical transmission chains where
in acoustic/articulatory space. learner acquires a language from the previous
selves each
learner then generates the data from which the next
This wide range of agent-based models suggest that key learner learns. A richer characterisation of learning was
design features of language emerge from iterated learn- provided by assuming that learners follow the principles
ing. Furthermore, the models employed by these authors of Bayesian inference, combining their own biases with
differ hugely in their approach (they include connection- the observed data (the linguistic behaviour of others)
ist models [27], exemplar models [28], grounded robotic when inferring a language. These biases, which capture
models [14], and induction of formal symbolic grammars the innate or acquired dispositions that make one
[18]), suggesting general principles at play in iterated language easier to learn than another, are expressed in
that transcend the particular model implementa- prior distribution over languages — a probability distri-
learning a
tion. bution where languages that are easier to learn are
assigned higher probability.
Mathematical models
The insights offered by agent-based simulations of iter- Griffiths and Kalish assumed that each learner made an
ated learning have recently been supplemented by math- inference by computing a posterior distribution over hy-
ematical results that characterise how languages can potheses that combined the biases reflected in the prior
change through cultural transmission. Mathematical mod- distribution with the information available in the linguis-
elling has been an important part of the theoretical de- tic data they had encountered. Each agent would then
velopment of evolutionary biology, and some of the tools choose a hypothesis by sampling from this distribution,
that have been developed for analysing biological evolu- and use this hypothesis to generate data for the next agent
prove equally powerful for analysing cultural evolu- a chain of transmission. Under these assumptions, the
tion in
tion. hypotheses selected by the agents converge to a particular
distribution as iterated learning proceeds: after enough
The potential of these mathematical tools was demon- episodes of transmission have passed, the probability that
strated in a series of papers by Nowak and colleagues a learner selects a particular hypothesis is just the prior
[29,30–32], who showed how one of the basic models of probability of that hypothesis, regardless of where the
biological evolution — the replicator dynamics — could process of iterated learning started.
be modified to capture aspects of language evolution. The
replicator dynamics indicates how the composition of a This convergence to the prior illustrates the potential power
of different types of organisms, each with a cultural transmission as an evolutionary force: even in
population of
different biological fitness, will change over time. By the absence of communicative interaction, iterated learn-
modifying this model to allow the fitness of each type ing can significantly change the languages spoken by a
to depend on the composition of the population, and for population. In particular, it can induce a shift towards
offspring to be of a different type from their parents, languages that are consistent with the biases of learners,
Nowak and colleagues were able to capture two important with those languages that are easiest to learn becoming
aspects of language evolution: the success of a language more prevalent in the population.
www.sciencedirect.com Current Opinion in Neurobiology 2014, 28:108–114
110 Communication and language
While this mathematical characterisation of iterated stimulus for a second participant, and so on. Bartlett
learning shares some of the conclusions of the agent- observed that material transmitted in this way changed
based simulation work reviewed above, in particular the as participants imposed their expectations about the
emphasis in some of the early work on the role of learner appropriate content onto the recalled material, causing
biases in shaping cultural evolution, there are also some it to be restructured: for instance, drawings might change
important mismatches. First, it indicated that there towards conventional representational forms (see also,
should be a one-to-one correspondence between the [41]). This is an experimental demonstration of the pre-
biases of learners and the extent to which a language is diction made by the mathematical analysis of iterated
likely to emerge through cultural transmission, while learning, outlined above, that systems of knowledge or
simulations had suggested that weak learning biases behaviour transmitted by iterated learning evolve to
could be magnified by iterated learning [18]. Second, reflect the biases of individuals involved in transmission.
iterated learning would result in convergence to the same
distribution — the prior — regardless of how much data Much of the modern work using the iterated learning
each learner saw. There was thus no effect of the learning paradigm with humans (see [42] for review) is of a similar
bottleneck in the mathematical analysis, in contrast to the nature, demonstrating the presence and consequence of
important role this seemed to play in simulations. learner biases. Several studies take known biases from
well-studied tasks, such as the learning of categories or
Attempting to reconcile these differences, Kirby, Dow- functions, and verify that transmission through iterated
man, and Griffiths [34] examined the effect of different learning yields behaviours which reflect those biases
learning mechanisms on the mathematical results. This [43,44]; an alternative approach is to use iterated learning
analysis showed that the differences from the simulation as a discovery procedure for biases, for example showing
were due to the assumption that learners sampled a in favour of retaining social information over non-
results biases
hypothesis from the posterior distribution. If learners social information [45], or using the results of iterated
adopt a more deterministic strategy — moving towards learning to arbitrate between theories of how people
simply selecting the hypothesis with highest posterior make predictions about everyday events [46].
probability — then iterated learning converges to a distri-
bution that exaggerates the prior: hypotheses with high In the domain of language evolution, several studies have
prior probabilities appear even more often, while those combined iterated learning techniques with artificial
with low prior probabilities become even less likely. The language learning or communication game paradigms
exact distribution depends on how much data is seen by to explore the way in which languages or other communi-
learner, with the prior having a stronger effect when systems evolve through learning and use (see [48]
each cation
only small amounts of data are available. This analysis for review). Kirby and colleagues introduced an iterated
thus helps to explain the circumstances under which learning paradigm (Figure 1) in which participants were
cultural transmission can magnify learning biases (allow- trained on an artificial language (a set of labels for
ing weak biases to be a potential explanation for strong coloured moving shapes) and then produced linguistic
linguistic patterns) and when a bottleneck effect will behaviour which subsequent individuals learnt from
emerge. Specifically, it suggests that future empirical [47 ] (see also [49]). A learning bottleneck was imposed
work should concentrate on the extent to which acqui- on transmission: while each participant produced a label
sition of language appears to involve sampling from a for the full set of stimuli, only a subset of those pictures
posterior or choosing the hypothesis that maximises the were presented, together with their labels, to the partici-
posterior. pant at the next generation. From an initial random
of objects, the languages changed over gener-
labelling
Subsequent mathematical analyses have begun to link ations so as to facilitate generalisation: as predicted by the
these results to broader questions about cultural and modelling results discussed above (e.g. [4]), compo-
biological evolution, exploring transmission in more com- sitional languages developed, where sub-parts of each
plex populations [35,36], the effects of the structure of the complex label specified components of the picture that
environment on the structure of language [37], formal label referred to (e.g. the first syllable of a complex label
relationships between iterated learning and the Wright- might indicate the colour, the second syllable might
Fisher model from population genetics [38], and the indicate shape, the third syllable movement). Related
biological evolution of learner biases [39]. experimental paradigms, in which participants learn or
communicate using a novel medium (systematically dis-
Laboratory experiments torted graphical scribbles, or a slide whistle) show the
The experimental study of iterated learning goes back at emergence of combinatorial structure, where complex
least as far as Bartlett’s ‘serial reproduction’ paradigm signals are composed by recombining a smaller set of
[40], in which participants were exposed to some stimulus meaningless component parts [50,51,52], again demon-
(e.g. a drawing), then asked to reproduce that material strating that the predictions of earlier agent-based mod-
from memory; their reproduced stimuli served as the elling above are borne out experimentally.
Current Opinion in Neurobiology 2014, 28:108–114 www.sciencedirect.com
Iterated learning and the evolution of language Kirby, Griffiths and Smith 111
Figure 1
Language 0 Language 1 Language 10
kalu lanapi nehoplo
Transmission Transmission
nane Set 0 kilahuna Set 1 nekiplo
pitu kalu Participant kinepilu Participant Participant nehopilu
1 kalu kilahuna 2 10
luki pilu nekipilu
mola neki kalu kinepilu lahoplo
kalakihu namola lamuna neki lakiplo
neki kinepila lahopilu
namola neki lakipilu
Current Opinion in Neurobiology
An illustration of the iterated artificial language learning method and indicative results, from [47 ]. Data shown is from their Experiment 2, Chain 3.
Participants are asked to learn a target language based on exposure to a subset of that language (labelled ‘Transmission Set’ here), with (a subset of)
the language produced by the nth participant in a chain of transmission providing the input to participant n + 1. In this experiment, participants were
asked to learn labels for coloured moving shape (there were 3 shapes, 3 colours, 3 motions: a subset are shown here). The initial language (Language
0) provided a randomly generated, idiosyncratic label for each such picture. As a result of the iterated learning procedure, this unstructured set of
meaning-signal associations developed into a structured language: in the chain shown here, by generation 10, each label consists of a prefix
specifying colour (e.g. ne- for black, la- for blue), a stem specifying shape (e.g. -ho- for circle, -ki- for triangle), and an affix specifying motion (e.g. -plo
for bouncing, -pilu for looping).
The combination of iterated learning and artificial and transmission chains, where drawings were transmitted
language learning has been used to show that miniature to naive individuals rather than within closed groups.
languages exhibiting unpredictable or ‘free’ variation These three population structures produce different types
absent from natural languages) become increas- graphical communication system. In dyads, participants’
(largely of
ingly regular and predictable as a result of their trans- drawings develop from rather complex affairs which
mission [53,54], demonstrating that adult learners have a represent their intended referent iconically (e.g. by resem-
bias in favour of regularity, and that these learning biases bling the actor or location they depict) to far simpler,
can explain the absence of unpredictable variation in economical but opaque symbols, which pick out their
natural languages (complementing studies which empha- intended referent only by convention within the dyad.
sise the role of strong biases in child learners in imposing In contrast, graphical representations in diffusion chains
regularity on language [55,56]). Using a similar exper- became increasingly complex and iconic. The systems
imental paradigm, [57] demonstrate that miniature voca- which emerge in communities differ more subtly from
bularies for describing colour evolve through iterated those which develop in dyads: community graphical
learning to resemble the distribution of colour naming representations are simple, like the representations that
systems observed in the world’s natural languages, again develop in dyads, but are less opaque to outsiders and
highlighting iterated learning as a mechanism which can inherently more ‘shareable’. Following on from this, other
explain linguistic universals. work in the same paradigm further explores the con-
sequences of transmission and interaction for the form
Other work has explored how the nature of interactions and structure of graphical communication systems [62–64].
between participants engaged in iterated learning can
shape an evolving communication system. In an important A range of iterated learning experiments have also been
series of studies [58–61], participants play a graphical carried out with non-human animals (see [65] for review),
communication game in pairs: the director produces a being primarily used to establish whether the studied
drawing which is intended to convey a concept to the species are capable of faithfully transmitting and main-
matcher, who attempts to identify the concept being con- taining a novel behaviour within a population [66 ]. In
veyed by the director. These studies compare simple dyads species where the presence of cultural transmission is
(two participants play together repeatedly), larger closed uncontroversial (e.g. songbirds), iterated learning has
groups of eight individuals (‘communities’), where mem- been used as a tool to investigate biases in learning, in
bers of the group play a series of pairwise communication close parallel to the experimental work with humans:
games, rotating through partners in a controlled fashion, Feher and colleagues show that an initially degenerate
www.sciencedirect.com Current Opinion in Neurobiology 2014, 28:108–114
no reviews yet
Please Login to review.
