Inductive learning is impossible without overhypotheses, or constraints on the hypotheses considered by the learner. Some of these overhypotheses must be innate, but we suggest that hierarchical Bayesian models help explain how the rest can be acquired. To illustrate this claim, we develop models that acquire two kinds of overhypotheses — overhypotheses about feature variability (e.g. the shape bias in word learning) and overhypotheses about the grouping of categories into ontological kinds like objects and substances.

Word learning is a ‘‘chicken and egg’’ problem. If a child could understand speakers ’ utterances, it would be easy to learn the meanings of individual words, and once a child knows what many words mean, it is easy to infer speakers’ intended meanings. To the beginning learner, however, both individual word meanings and speakers ’ intentions are unknown. We describe a computational model of word learning that solves these two inference problems in parallel, rather than relying exclusively on either the inferred meanings of utterances or cross-situational word-meaning associations. We tested our model using annotated corpus data and found that it inferred pairings between words and object concepts with higher precision than comparison models. Moreover, as the result of making probabilistic inferences about speakers’ intentions, our model explains a variety of behavioral phenomena described in the word-learning literature. These phenomena include mutual exclusivity, one-trial learning, cross-situational learning, the role of words in object individuation, and the use of inferred intentions to disambiguate reference.

What is the relation between the material, conventional symbol structures that we encounter in the spoken and written word, and human thought? A common assumption, that structures a wide variety of otherwise competing views, is that the way in which 10 these material, conventional symbol-structures do their work is by being translated into some kind of content-matching inner code. One alternative to this view is the tempting but thoroughly elusive idea that we somehow think in some natural language (such as English). In the present treatment I explore a third option, which I shall call the ‘‘complementarity’ ’ view of language. According to this third view the actual symbol 15 structures of a given language add cognitive value by complementing (without being replicated by) the more basic modes of operation and representation endemic to the biological brain. The ‘‘cognitive bonus’ ’ that language brings is, on this model, not to be cashed out either via the ultimately mysterious notion of ‘‘thinking in a given natural language’ ’ or via some process of exhaustive translation into another inner code. Instead, 20 we should try to think in terms of a kind of coordination dynamics in which the forms and structures of a language qua material symbol system play a key and irreducible role. Understanding language as a complementary cognitive resource is, I argue, an important part of the much larger project (sometimes glossed in terms of the ‘‘extended mind’’) of understanding human cognition as essentially and multiply hybrid: as involving 25 a complex interplay between internal biological resources and external non-biological resources.

thank members of the UBC Baby Cognition Lab for their help with data collection, and Paul Bloom, Geoff Hall, and Terry Regier for helpful discussion. We owe a particular debt to Liz Bonawitz, for discussions and pilot work on an earlier version of this work.

Abstract not found

This article is a response to E. Heit and B. K. Hayes’s (2005) comment on the target article “Induction

Young children often exhibit flexible behaviors relying on different kinds of information in different situations. This flexibility has been traditionally attributed to conceptual knowledge. Reported research demonstrates that flexibility can be acquired implicitly and it does not require conceptual knowledge. In Experiment 1, 4- to 5-yearolds successfully learned different context-predictor contingencies and subsequently flexibly relied on different predictors in different contexts. Experiments 2A and 2B indicated that flexible generalization stems from implicit attentional learning rather than from rule discovery, and Experiment 3 pointed to very limited strategic control over generalization behaviors in 4- to 5-year-olds. These findings indicate that mundane mechanisms grounded in associative and attentional learning may give rise to smart flexible behaviors. Even early in development, people’s generalization is remarkably flexible—depending on a situation, people may rely on different kinds of information. This flexibility has been found in a variety of generalization tasks, including lexical extension, categorization, and property induction. For example, in a lexical extension task (Jones, Smith, & Landau, 1991), 2- to 3-year-olds were presented with a target, which was named (i.e., ‘‘this is a dax’’), and asked to find another dax among test items. Children extended the label by shape alone when the target and test objects were presented without eyes. However, they extended the label by shape and texture when the objects were presented with eyes. Children exhibit similar flexibility in categorization and induction tasks. For example, in a categorization task, 3- to 4-year-olds were more likely to group items on the basis of color if the items were introduced as food, but on the basis of shape if the items were introduced as toys (Macario, 1991). In another task, 4- to 5-year-olds were presented with a target and two test items, such that one test item shared the label with the target and the other looked similar to the target. Participants were then told that the target had a particular property and asked which

In typical development, word learning goes from slow and laborious to fast and seemingly effortless. Typically developing 2-year-olds are so skilled at learning noun categories that they seem to intuit the whole range of things in the category from hearing a single instance named – they are biased learners. This is not the case for children below the 20th percentile on productive vocabulary (late talkers). This paper looks at the vocabulary composition of age-matched 18-30-month-old late- and early-talking children. The results of Experiment 1 show that late talkers ’ vocabularies are more variable than early talker’s vocabularies. Crucially, Experiment 2 shows that neural networks trained on the vocabularies of individual late talkers learn qualitatively different biases than those trained on early talker vocabularies. These simulations make testable predictions for world learning biases of late- vs. early-talking children. The implications for diagnosis and intervention are discussed.

Inductive learning is impossible without overhypotheses, or constraints on the hypotheses considered by the learner. Some of these overhypotheses must be innate, but we suggest that hierarchical Bayesian models help explain how the rest can be acquired. The hierarchical approach also addresses a common question about Bayesian models of cognition: where do the priors come from? To illustrate our claims, we consider two specific kinds of overhypotheses — overhypotheses about feature variability (e.g. the shape bias in word learning) and overhypotheses about the grouping of categories into ontological kinds like objects and substances. Compared to machine-learning algorithms, humans are remarkable for doing so much with so little. A single

All students of cognitive development agree that the central questions in development are 1) specifying the initial state of a human infant, 2) specifying the final state of development for a human adult, and 3) specifying how to get from the initial state to the final state. Then academic disputes ensue. Cognitive developmental psychologists are roughly divided into two camps: those who are more or less nativists and those who are more or less empiricists. Some psychologists do not like these terms, and some alternatives are “those who believe in innate knowledge ” and “those who believe in learning, ” or “those who believed in initial conceptual knowledge ” and “those who believe in initial perceptual capabilities. ” This division is also correlated with whether a researcher believes in domain specificity or not: nativists tend to argue for domain-specific knowledge (even at the beginning of development) and domain-specific learning mechanisms; empiricists tend to argue for domain-general learning mechanisms that may result in domain-specific knowledge some years into development (for some representative explications of these views, see Carey &