Language learning is an iterative process, with each learner learning from other learners. Analysis of this process of iterated learning with chains of Bayesian agents, each of whom learns from one agent and teaches the next, shows that it converges to a distribution over languages that reflects the inductive biases of the learners. However, if agents are taught by multiple members of the previous generation, who potentially speak different languages, then a single language quickly dominates the population. In this work, we consider a setting where agents learn from multiple teachers, but are allowed to learn multiple languages. We show that if agents have a sufficiently strong expectation that multiple languages are being spoken, we reproduce the effects of inductive biases on the outcome of iterated learning seen with chains of agents. 1.

We simulate the evolution of a domain language in small speaker communities. Data from experiments (Garrod et al., 2007; Fay et al., 2008) show that human communicators can evolve graphical languages quickly in a constrained task (Pictionary), and that communities converge towards a common language even in the absence of feedback about the success of each communication. We postulate that simulations of such horizontal evolution have to take into account properties of human memory (cue-based retrieval, learning, decay). We implement a model that can draw abstract concepts through sets of non-abstract, related concepts, and recognize such drawings. The knowledge base is a network with association strengths randomly sampled from a natural distribution found in a text corpus; it is a mixture of knowledge shared between agents and individual knowledge. In three experiments, we show that the agent communities converge, but that initial convergence is stronger when communities are structured so that the same pairs of agents interact throughout. Convergence is weaker in communities when agents do not swap roles (between recognizing and drawing), predicting the necessity of bi-directional communication in domain language evolution. Average and ultimate recognition performance depends on how much of the knowledge agents share initially.

Abstract not found

Diversity of language is a key part of our understanding of natural languages now and from the past. This diversity goes hand in hand with language change. Change is pervasive at every linguistic level. However, the space of existing languages does not appear to be unconstrained. In the modern generative tradition, this is governed by Universal Grammar (UG) (see, for example, Kroch (2000)).

Recent research on computational models of language change and cultural evolution in general has focused on the analytical study of languages as dynamic systems, thus avoiding the difficulties of analysing the complex multi-agent interactions underlying numerical simulations of cultural transmission. The same is true for the examination of the effects of inductive biases on language distributions within the Bayesian Iterated Learning Framework. The aim of this work is to test whether the strong results obtained through analytical methods in this framework also extend to finite populations of Bayesian learners, and to investigate what other effects richer population dynamics have on the results. Small world networks are introduced as a tool to model social structures which are shown to play an important role in the outcome of cultural transmission processes. The assumptions behind a Bayesian approach to language learning and its implications will be studied and compared to previous models of language change. While studying the effects of populations on convergence rates in the Bayesian model, the role of more complex population settings for the future of Iterated

We simulate the evolution of a domain vocabulary in small communities. Empirical data show that human communicators can evolve graphical languages quickly in a constrained task (Pictionary), and that communities converge towards a common language. We propose that simulations of such cultural evolution incorporate properties of human memory (cue-based retrieval, learning, decay). A cognitive model is described that encodes abstract concepts with small sets of concrete, related concepts (directing), and that also decodes such signs (matching). Learning captures conventionalized signs. Relatedness of concepts is characterized by a mixture of shared and individual knowledge, which we sample from a text corpus. Simulations show vocabulary convergence of agent communities of varied structure, but idiosyncrasy in vocabularies of each dyad of models. Convergence is weakened when agents do not alternate between encoding and decoding, predicting the necessity of bi-directional communication. Convergence is improved by explicit feedback about communicative success. We hypothesize that humans seek out subtle clues to gauge success in order to guide their vocabulary acquisition.

interaction?

Human languages vary in many ways, but also show striking cross-linguistic universals. Why do these universals exist? Recent theoretical results demonstrate that Bayesian learners transmitting language to each other through iterated learning will converge on a distribution of languages that depends only on their prior biases about language and the quantity of data transmitted at each point; the structure of the world being communicated about plays no role (Griffiths & Kalish, 2005, 2007). We revisit these findings and show that when certain assumptions about the independence of languages and the world are abandoned, learners will converge to languages that depend on the structure of the world as well as their prior biases. These theoretical results are supported with a series of experiments showing that when human learners acquire language through iterated learning, the ultimate structure of those languages is shaped by the structure of the meanings to be communicated.

Bayesian models of human causal induction rely on assumptions about people’s priors that have not been extensively tested. We empirically estimated human priors on the strength of causal relationships using iterated learning, an experimental method where people make inferences from data generated based on their own responses in previous trials. This method produced a prior on causal strength that was quite different from priors previously proposed in the literature on causal induction. The predictions of Bayesian models using different priors were then compared against human judgments of strength of causal relationships. The empirical priors estimated via iterated learning resulted in the best predictions.

Progress in studying human categorization has typically involved comparing generalization judgments made by people to those made by models for a variety of training conditions. In this paper, we explore an alternative method for understanding human category learning—iterated learning—which can directly expose the inductive biases of human learners and categorization models. Using a variety of stimulus sets, we compare the results of iterated learning experiments with human learners to results from two prominent classes of computational models: prototype models and exemplar models. Our results indicate that human learning is not perfectly captured by either type of model, lending support to the theory that people use intermediate representations between these two extremes.