There is a deep theoretical tension at the heart of cognitive science. Human beings have abstract, hierarchical, structured, and accurate representations of the world: representations that allow them to make wide-ranging and correct predictions. They also learn those representations. They derive them from concrete, particular, and probabilistic combinations of experiences. But how can we learn abstract structure from the flutter and buzz at our retinas and eardrums? Nativists, from Plato to “core cognition” theorists, argue that it only seems that we learn; in fact, the abstract structure is innate. Empiricists, from Aristotle to connectionists, argue that it only seems that we have abstract structure; in fact, we just accumulate specific sensory associations. When we see both abstract structure and learning – notably in scientific theory change – traditional nativists and empiricists both reply that such conceptual change requires elaborate social institutions and explicit external representations.

Carey has made major contributions to the enormous empirical progress of cognitive development (Carey Reference Carey2009). But those very empirical discoveries have actually made the conceptual problem worse. Piaget could believe that children started out with specific sensorimotor schemes and then transformed those schemes into the adult's abstract representations. But Carey's own studies, along with those of others, have shown that this is not a feasible option. On the one hand, contra the empiricists, even infants have abstract structured knowledge. On the other hand, contra the nativists, conceptual theory change based on experience takes place even in childhood, without the infrastructure of adult science.

This book tries to resolve that conceptual tension by a sort of division of labor between nativism and empiricism. Cognition in infancy and early childhood reflects “core cognition” – the nativist option. Conceptual change is a later development that operates on the representations of core cognition but requires language, and the somewhat mysterious process of “Quinian bootstrapping.” However, new empirical and computational work, much of it done only in the past few years, suggests that there are more coherent ways of solving this dilemma. We need not propose a radical discontinuity between the processes that are responsible for abstract structure and those that are responsible for learning and conceptual change. They are, in fact, both part of the same system, and they are in place from the beginning.

Empirically, we've discovered that even infants have powerful learning capacities (Woodward & Needham Reference Woodward and Needham2009). We can give children particular types of evidence and observe the types of structure that they induce. These studies have already shown that infants can detect complex statistical patterns. But, more recently, it has been discovered that infants can actually use those statistics to infer more abstract non-obvious structure. For example, infants can use a statistically nonrandom pattern to infer someone else's desires (Kushnir et al. Reference Kushnir, Xu and Wellman2010), and can use statistical regularities to infer meanings (Graf Estes et al. Reference Graf Estes, Alibali, Evans and Saffran2007; Lany & Saffran Reference Lany and Saffran2010). In other experiments, giving infants relevant experience produces novel inferences both in intuitive psychology and in intuitive physics (Meltzoff & Brooks Reference Meltzoff and Brooks2008; Somerville et al. Reference Somerville, Woodward and Needham2005; Wang & Baillargeon Reference Wang and Baillargeon2008).

By the time children are 4 years of age there is consistent evidence both for conceptual change and for learning mechanisms that produce such change. Pressing children to explain anomalous behavior can induce a representational understanding of the mind (Wellman & Liu Reference Wellman, Liu, Gopnik and Schulz2007), and giving them a goldfish to care for can provoke conceptual changes in intuitive biology (Inagaki & Hatano Reference Inagaki and Hatano2004). Most significantly, preschoolers can use both statistical patterns and active experimentation to uncover complex and abstract causal structure, inducing unobserved causal forces and high-level causal generalizations (Gopnik et al. Reference Gopnik, Glymour, Sobel, Schulz, Kushnir and Danks2004; Lucas et al. Reference Lucas, Gopnik, Griffiths, Ohlsson and Catrambone2010; Schulz & Bonawitz Reference Schulz and Bonawitz2007; Schulz et al. 2007; Reference Schulz, Goodman, Tenenbaum and Jenkins2008). Empirically, even infants and very young children seem to use statistical inference, explanation, and experimentation to infer abstract structure in a way that goes well beyond association and could support conceptual change.

We can still ask how and even whether this sort of learning is possible computationally. Fortunately, new work in the “probabilistic models” framework, both in cognitive development and in the philosophy of science and machine learning, provides a promising answer (Gopnik & Schulz Reference Gopnik and Schulz2007; Gopnik et al. Reference Gopnik, Glymour, Sobel, Schulz, Kushnir and Danks2004; Griffiths et al. Reference Griffiths, Chater, Kemp, Perfors and Tenenbaum2010; Xu & Tenenbaum Reference Xu and Tenenbaum2007). On this view, from the very beginning, cognition involves the formulation and testing of abstract hypotheses about the world, and, from the very beginning, it is possible to revise those hypotheses in a rational way based on evidence.

The new idea is to formally integrate structured hypotheses, such as grammars, hierarchies, or causal networks, with probabilistic learning techniques, such as Bayesian inference. The view is that children implicitly consider many hypotheses and gradually update and revise the probability of those hypotheses in the light of new evidence. Very recently, researchers have begun to show how to use these methods to move from one abstract high-level framework theory to another: the sort of conceptual change that Carey first identified (Goodman et al. Reference Goodman, Ullman and Tenenbaum2011; Griffths & Tenenbaum 2007). Empirically, we can induce such change, producing, for example, a new trait theory of actions (Seiver et al. 2010). Of course, there is still a great deal of work to be done. In particular, we need more realistic accounts of how children search through large hypothesis spaces to converge on the most likely options.

The new empirical work and computational ideas suggest a solution to Carey's dilemma – one that does not require either core cognition as a vehicle for abstract structure, or language and analogy as agents of conceptual change. It is also quite possible, of course, that the balance of initial structure, inferential mechanisms, and explicit representational resources might differ in different domains. Mathematical knowledge, is, after all, very different from other types of knowledge, ontologically as well as epistemologically, and might well require different resources than spatial, causal, or psychological knowledge.

In general, however, there is real hope that the empirical work and theoretical ideas that Carey has contributed can be realized in an even deeper way in the new computational theories, and that the ancient tension she has elucidated so well can finally be resolved.