Carey presents a theory of the ontogenesis of concepts marked by the presence of innate core knowledge and the representation of causality on one side, and by Quinian linguistic bootstrapping on the other (Carey Reference Carey2009). We endorse most of this view. We find the evidence for the core knowledge of agents and objects convincing, and we agree that concepts such as electron are learned by acquiring linguistic placeholder structures first. But what happens in-between these two developmental landmarks? Is there any genuine conceptual development, including forming new kind-sortals with theory-constrained representation of features, apart from linguistic bootstrapping? How is the integrative role of causal representations (Carey Reference Carey2009, Ch. 6) being accomplished?

Something of a paradox emerges when one seeks evidence of such processes in infants less than 12 months of age. A massive body of research from the last 25 years documents impressive innate competencies and capacity for rapid conceptual development over first months and years of life (Carey Reference Carey2009, Ch. 3–6). However, another body of research shows that early representations are in many aspects impoverished, and that their developmental trajectories sometime contain surprising gaps. One of the most striking is perhaps a decalage in using featural, kind-related information for object identification and object individuation (Carey Reference Carey2009, Ch. 3; Xu Reference Xu2007b). On the one hand, even 6-month-olds appreciate that when there are two differently colored objects in the launching event, these are two separate objects, and infants show surprise when one takes the other's causal role (Carey Reference Carey2009, Ch. 6; Leslie Reference Leslie1982). But on the other hand, infants less than 12 months of age are not concerned either with shape, color, or kind, when determining number of occluded objects (Carey Reference Carey2009, Ch. 3; Xu & Carey Reference Xu and Carey1996). It seems that infants use two different, separate types of object representations, and that in their first year they do not form sortal categories representing permanent object-properties, that is, kind concepts. There is evidence for sortals object and agent in early infancy (Surian & Caldi Reference Surian and Caldi2010), but these could be the only sortals available during the very first months.

However, some recent results suggest more permanent, feature-based representations in the first year of life, and relatively complex representations of this kind only a few months later. A series of studies by Wilcox and others (McCurry et al. Reference McCurry, Wilcox and Woods2008; Wilcox et al. Reference Wilcox, Woods, Toggy and Napoli2006) demonstrated use of featural information for determining number of objects in the event even at 4.5 months. Featural representations made by 5- and 3-month-olds support success in goal-attribution tasks (Luo Reference Luo2010; Schlottmann & Roy 2009). Moreover, infants between 3.5 and 10 months studied by Hamlin et al. (Reference Hamlin, Wynn and Bloom2007; Reference Hamlin, Wynn and Bloom2010) map permanent (at least in the experimental session's time-scale) dispositional states on feature-based object representations derived from one event (different colored shapes helping or hindering one another), and use this representation either to apprehend shapes' behavior or to select an object to reach for during a different test-event. Nine-month-olds use property-to-function mapping derived form one set of objects for individuation within another set (Wilcox et al. Reference Wilcox, Woods and Chapa2009), Early in the second year, children represent occluded colors to predict others' behavior (Luo & Beck Reference Luo and Beck2010) and, as our own work in progress suggests, represent feature sequences of an object undergoing self-induced transformation (Hernik & Haman Reference Hernik and Haman2010). However, as most of these studies either do not test directly for numerical identity processing, or use simplified procedures, or objects-to-be-represented are involved in only a single event, it can still be argued that “infants represent objects relative to the goals of agents” (Carey Reference Carey2009, p. 450) and feature-based kind-sortals remain a relatively late developmental achievement.

Other explanations are also possible, however (Xu Reference Xu2007b). Wang and Baillargeon developed a research program aiming at demonstrating that at least in the physical domain, infants' reliance on event-representation leads to featural representations of objects and arguably to proto-representations of kinds (see Wang & Baillargeon Reference Wang and Baillargeon2009 for review). It can be presented as a three-way bootstrapping model where: (1) object-file representations, restricted by attention span, provide spatiotemporal cues for object individuation and placeholders for features assigned both bottom-up and top-down; (2) event category representations provide causal relations and guide attention to causally-relevant features; and (3) feature-based object representations by means of which temporary information provided by the aforementioned two systems is converted into permanent mental structures. Context-independence of these representations increases with development, and can be successfully trained (Wang Reference Wang2011).

Although Wang and Baillargeon (Reference Wang and Baillargeon2009) present their model in the context of early physical knowledge, we are tempted to conceive it is an example of complex bootstrapping potentially more common among mechanisms of conceptual development. The structures of causal knowledge involved in different categories of physical and social events, functional representations, or perhaps object transformations, as in the studies mentioned earlier, provide scaffolding for feature- and kind-representations. Different kinds of representations are used, and each of them provides placeholders, which can bootstrap elements of other systems of representation. This multiway bootstrapping may be further supported by at least three general-purpose learning mechanisms, such as weak linguistic influence (Carey Reference Carey2009, Ch. 7; Xu Reference Xu and Gaskell2007a) and learning by means of communication (Csibra & Gergely Reference Csibra and Gergely2009; Futo et al. Reference Futo, Teglas, Csibra and Gergely2010), which may support both kind- and feature-based representations, as well as constrained Bayesian learning, which allows for quick detection of conditional interrelations between causes, objects, and their features. So far this is only a speculative hypothesis. Although integration of mechanisms discussed here can in principle be explored in experimental designs, it rarely was. Most of the contemporary cognitive-developmental research aimed at delimitating and enumerating separate mechanisms. Whereas this strategy was justified and effective, it can be perhaps blamed for underestimating the developmental role of integrative processes, and in consequence underestimating infants' early representational potential. Now, the focus on cooperation of different systems in the first years of life is necessary (see Denison & Xu Reference Denison and Xu2010 for a convergent argument focusing on Bayesian learning). We suppose that a more integrative research program could fill the gap in our understanding of how everyday concepts begin to be formed very early on, in the midst of capitalizing on two major developmental forces (initial knowledge and linguistic bootstrapping) described by Carey.