R2. Advances in interactive specialization

Innocenti and Anderson both make specific comments about processes related to the Interactive Specialization model (Johnson Reference Johnson2000; Reference Johnson2001) presented within our discussion of embrainment. Anderson presents a method for establishing the degree of unique specialization of cortical regions, defined as a cortical region that is involved in performing some computations but not others. In particular, he argues for a distinction between computational specialization and domain specialization. In computational specialization, regions develop increasingly well-defined subfunctional roles within an activated network; with respect to cognitive domains, specialization would be a property of the particular network of regions. In domain specialization, it is the regions themselves that become increasingly dedicated to a narrow range of cognitive functions, thereby restricting their involvement in other domains. Anderson then presents initial findings from a meta-analysis of functional magnetic resonance imaging (fMRI) studies and argues that they provide support for the computational specialization view: The unit of domain specialization appears to be the network of regions rather than regions per se.

While we have not defined the end-state of Interactive Specialization in the level of detail described by Anderson, we agree with his conclusions that the results of specialization may be cortical regions that play multiple roles in different domains. Each region has an increasingly specialized role as part of the network activated to support the computations necessary for a given domain. Indeed, this notion resonates well with the general idea of partial representations. However, developmental work argues that active brain areas move from broad and fuzzy, to less broad and more specialized (e.g., Durston et al. Reference Durston, Davidson, Tottenham, Galvan, Spicer, Fossella and Casey2006). Such a shift would seem to indicate that regions lose at least some of their contribution to specific cognitive functions across development.

Anderson's approach is simple but powerful and offers provocative results. Of course, there are limitations to what a meta-analysis can achieve, imposed in the main by the quality and consistency of the contributing studies. The practical limitations of fMRI studies include the predominant reliance on subtraction methodologies, the poor temporal acuity, the mix of whole-brain analyses and region of interest analyses, and the variability of techniques employed within the analyses themselves (smoothing functions, thresholds, etc.). Some of the apparent involvement of multiple cortical areas in multiple tasks may be a consequence of the limitations of (some of) the studies included in the meta-analysis.

In our own work, we have embraced the distinction that Anderson has identified, and designed fMRI paradigms to test it empirically. The most important requirement, here, is that the cortical areas activated by different tasks should be evaluated within the same participant; otherwise between-participant variability may be mistaken for distributed processing. That is, if a region performs function x in one participant but function y in another participant, one might (mistakenly) infer that this region performs both function x and function y in the normal human brain. In the lab of one of our authors (M.S.C.T.), one current study employs multiple tasks within the broader domain of language (auditory comprehension, reading, naming, semantic retrieval, and gesture production) carried out by the same participants. The analyses target four regions of interest to examine their relative involvement in the functional networks activated during each task. The design is event-related to allow for analyses of functional connectivity between the regions; and the study uses a lifespan design with a cross-sectional sample from children of 7 years old up to adults of 75, to explore whether regional or network specialization alters with development and/or aging. The practical limitations of a paradigm such as this for addressing interactive specialization are that many of the interesting developmental processes will have been (largely) completed by the youngest ages in which it is practical to run the paradigm; and even with an event-related design, the temporal acuity can offer only a very blurred view of the computational relations between the regions involved. fMRI designs must therefore be complemented by methods that are more readily applicable to infancy and that offer better temporal acuity, such as electrophysiology (see, e.g., De Haan Reference De Haan2007; Nelson & Monk Reference Nelson, Monk, Nelson and Luciana2001).

Additionally, we suspect there is further progress to be made in exploring interactive specialization through computational modeling. Within this discipline, one must precisely specify what it would mean for emergent specialization to be (in the main) a property of networks of regions, with each region participating in many such networks. What exactly is the computational role of a region in these networks? Does the region offer, for example, modality-specific partial representations (e.g., dimensions of visual information)? Or a particular computational primitive (e.g., inhibitory control, analysis of sequences)? Or is it a chameleon, adopting different computational guises in each functional network? For ease of explanation, we would probably prefer the first or second options to be true. However, even simple connectionist networks can learn to demonstrate the third property.Footnote ¹ Such “context sensitive” computations are intrinsic to connectionist networks, which is why we believe they are an excellent tool to investigate cognition within a neuroconstructivist framework. The key issue would then become to understand how such chameleonlike computational properties are exploited by, and change over, development.

R3. Neuroconstructivism versus selectionism

Innocenti argues that selectionism, the process of pruning neural connections to cull those that are irrelevant, noisy, or redundant, may provide a basis for Interactive Specialization. We agree wholeheartedly, with the caveat that constructive neural processes may also play an important role. It is worth making clear the perceived tension between constructivism and selectionism, before indicating why we think the two are in fact largely consistent with each other. Constructivist development argues for a progressive increase in the complexity of representations, which enables new competences to develop based on earlier, simpler ones. Within neuroconstructivism, this increase in representational complexity is realized in the brain by a progressive elaboration of functional cortical structures. By contrast, Innocenti points to the large accumulation of empirical evidence for the generalized occurrence of exuberance and regression in brain development. Under the selectionist view “adult neural (and … cognitive structures) emerge by selection from a richer (but constrained) juvenile stock” as Innocenti puts it. On the one hand we have elaboration while on the other we have regression.

We believe these positions are consistent if one views the exuberance and regression in terms of (potential) computational resources, and the elaboration of structures in terms of actual functions. For us, selection is an outcome of the fundamental processes of cooperation and competition that lead to the emergence of partial representations. Innocenti seems agnostic as far as how goal-directed the exuberance preceding the regression is, but a fair amount is now known about the experience-dependent, goal-directed elaboration of neural structures (see, e.g., Quartz & Sejnowski Reference Quartz and Sejnowski1997). From a constructivist viewpoint, there is also a sense in which the emergence of partial representations in different regions sets the stage for the emergence of specialized networks of regions, to deliver more complex cognitive functions (as discussed above).

Shultz outlines a novel type of constructivist neural network that captures this idea. In knowledge-based cascade correlation (KBCC), the overall system tests and recruits networks from existing stock when they are required to support new computations. We speculate that a KBCC network architecture may capture, at least to some extent, emerging functionality in the prefrontal cortex (PFC). Converging lines of evidence show that the PFC has an important role in orchestrating other cortical regions during functional brain development (Johnson Reference Johnson2005; Thatcher Reference Thatcher1992), as Foster, van Eekelen, & Mattes (Foster et al.) point out. We believe that constructivist neural networks of the kind described by Shultz are worth exploring as models of PFC functional development, in addition to the selective pruning process preferred by Innocenti. This leads us to a wider consideration of the nature of representations within the neuroconstructivist framework.

R4. Representation: Multiple modalities, flexibility, and abstraction

Our emphasis on the emergence of partial representations to drive behavior elicited several kinds of responses. Bremner & Spence argue for the importance of multisensory/multimodal experiences to deliver enriched representations that serve more than the child's immediate behavioral goals. French applauds the combination of dynamic systems theory and connectionist approaches inherent in the notion of partial representations. His research in analogy-making has led him to endorse the same idea (or “gospel”) of flexible, context-dependent representations that are gradually constructed via a continual interaction between high- and low-level constraints. Kotchoubey argues that the very concept of representation implies a Cartesian theatre and the need for a homunculus to whom the content of representations is made available. Moreover, he wonders how “full” instead of partial representations might ever come about. For Campos & Sotillo, the emergence of specialized, context-dependent representations must be complemented by the later emergence of explicit representations that “permit independence from context, conscious reasoning, and interrepresentational relations,” in line with the theory of representational redescription proposed by Karmiloff-Smith (Reference Karmiloff-Smith1992). Indeed, the notion that context-dependent partial representations are the outcome of interactive specialization prompts further questions. For example, how does this square with the idea of the increasing abstraction of the child's representation over development? How does it fit with the idea of increasing flexibility in the child's use of knowledge, for example in the case of the development of meta-cognition? If anything, is this not an increase in context independence?

First, we agree with Bremner & Spence regarding the importance of multisensory and multimodal representations. We acknowledge that we may not have emphasized this sufficiently: Multimodal integration is a major constraint on development that played an important role in Piaget's theories and has been vindicated by neuroscientific research. We expect that in many cases, partial representations will be the consequence of detecting contingencies between multiple sets of sensory and motor information in regions where this information converges, and that richer information will aid in the detection of those contingencies. When a region receives multiple inputs, this is termed a fan-in. We also expect fan-out to be important – that is, where a region drives several outputs (see, e.g., the model of infant-object interactions described in chapt. 9 of Vol. I). Indeed, to the extent that the brain exhibits bidirectional connectivity, multiple outputs may be as influential as multiple inputs in driving the development of the representations in a given region.

Returning to our framework, we see multisensory integration as falling under the principle of embrainment, where the development and processing in one region is influenced by different brain regions (in this case, sensory cortices). This is similar to the sensorimotor model of the acquisition of speech sounds described in chapter 8 of Volume I. The same model has indeed been used to account for effects of multisensory integration (Westermann Reference Westermann, Moore and Stenning2001). In this model two neural maps were linked via developing Hebbian connections. By exposing the model to multidimensional data sets with different correlational structures, Westermann demonstrated how perceptual change in each domain (such as categorical perception and perceptual prototypes) as well as multidimensional integration (the McGurk effect) can emerge from the gradual integration of unimodal data into a unified multimodal percept.

There is, however, a sense in which the idea of multisensory integration depends on the level of description. In general, we would expect the same neurocomputational mechanisms to operate between domains as within a single domain. Forming a (partial) representation requires learning about the statistics of patterns of presynaptic activity. The neuron doing the learning cannot know if the inputs it is receiving come from one or multiple domains. So the mechanism used must be the same in both cases. Multisensory integration is therefore a property that is salient at a particular (architectural) level of description.

Are the representations that result from the convergence of multiple senses any more abstract than those that result from a single sense? Possibly, in the sense that they are not bound to a single modality alone. A further sense in which increases in abstraction may be the consequence of development is the detection of higher-order invariances found in progressively higher layers of self-organizing neural networks, for example as observed in the mammalian visual system (see review in Ch. 6 of Vol. I). A more challenging prospect is to explain increases in flexibility in terms of context-dependent partial representations. We suspect the explanation will be in terms of the development of executive functioning, where flexibility is the ability to take advantage of existing knowledge in driving new behaviors. We are in agreement with Foster et al. on this point, as discussed in the previous section. Does the idea of flexibility necessarily imply context independence? We believe it does not. Equally, it may constitute the ability to switch back and forth between different contexts in a system whose representations remain context-bound.

Turning to the concerns of Kotchoubey, we do not agree that the concept of a representation or of an informational state necessarily requires the presence of an observer to read off the contents of the representation. The modified synapse is modified, whether anyone notices or not. Perhaps more important, it is not clear what empirical implications bear on this issue, one way or another. What consequences does it have for the developmental cognitive neuroscientist studying the mind whether representations require an observer or not? What consequences does it have for the developmental roboticist whose robot continues to show increasingly adaptive behavioral contingencies in interacting with its environment, even after its programmer has left the room?

In response to the idea of a partial representation, Kotchoubey justifiably poses the question of how an individual might ever attain a “full” representation of a concept. Here, we agree with Innocenti when he argues that “partial representations seem to be economical in terms of neural resources and are open to completion by developmental experience and/or learning, in adult life.” We claim the following: For a representation to become full, the individual must integrate the partial representations across the entire range of contexts in which the concept is used. Different subsets of partial representations may be appropriate for different contexts and indeed, the only feature common to all contexts of the full representation may be the presence of a unifying label. In a sense, the generation of a full representation is similar to the process undertaken by a philosopher when he or she considers the coherence of a concept. What are all the applications of an idea? Philosophers, however, usually add a further proviso that concepts should not be context sensitive – a concept should mean same thing in all situations. While this proviso may optimize philosophical discourse, it may not be a principle to which the brain adheres.

Lastly, both Bremner & Spence and Campos & Sotillo raise the important issue of the outcome of development. The former argue that representational states are richer than those simply required to drive behavior in the moment, while the latter point to the importance of the late-developing property of explicitization that enables performance to break free from the context of its acquisition. In each case, development might be argued to have gone “beyond mastery” (Karmiloff-Smith Reference Karmiloff-Smith1992). We agree that there are key issues to be explored here. To these challenges, we might add behaviors that appear decoupled from the immediate environment, such as imagination and play. However, we would argue that all of these are amenable to the neuroconstructivist approach, with its appeal to multidisciplinary sources of evidence and formal specification of causal mechanism. We feel that the important ideas of Rivière, as discussed by Campos & Sotillo, may be in particular need of formal specification before they can be investigated at a neurocognitive level. For example, is explicitization a process intrinsic to a representation, where a system somehow self-organizes into a structure that generates new properties in the absence of a behavioral driver? Or is it more like an enabling of connectivity across systems, perhaps in particular to the language systems that drive (explicit) verbal report? When a child moves beyond mastery, is this more than a continued proactive tendency to query the environment for further potential sources of reward (novelty) – an idea well understood within the framework of reinforcement learning? Or does it refer to an internal computational process, for example a system-wide preference for global consistency across components of an interactive processing system, beyond the local consistency that enables a given subsystem to generate a behavior successfully?

R5. Genetics, lifespan, and evolution

French argues that we have given short shrift to the role of genetics in cognition, and views as “bordering on the incomprehensible” our statement that “in humans, genes provide only very broad sorts of constraints on the representations that emerge in cortex.” We are sympathetic to his view; it is true that apart from a six-page section in chapter 1 of Volume I where we discuss the genetics of development, genes only appear intermittently throughout the volume. We have endeavored to rectify this in other publications (see Westermann et al. Reference Westermann, Mareschal, Johnson, Sirois, Spratling and Thomas2007, Fig. 4) where we emphasize the place of genes in our framework, and in particular the probabilistic epigenesis view of development (Gottlieb Reference Gottlieb1992; Reference Gottlieb2007). In the latter view, environmental and behavioral influences play a fundamental role in modulating the expression of genes. We expect that genes will exhibit a reliance on context (of other genes and of environments) for their causal effects similar to the contextual effects we observe at higher levels of description. We might term the contextual effect at the molecular level as engenement.

Despite our sympathies, there is also a sense in which we would defend our stance, and hold to our claim regarding the influence of genes on the emergence of representations in the cortex. While genes inevitably play an essential role in constructing and maintaining biological organisms from the moment of the fertilization of the egg onwards, the key issue at stake is the specificity of their influence on cognition. At present, while there are many suggestive correlations between gene variants or mutations and cognitive variability, there are few examples of explicit causal pathways that link these levels of description. We maintain that the optimum level of description to explain children's behavior is the cognitive level; the aim of neuroconstructivism is to seek constraints from other levels of description but not to reduce to them. And the causal chains linking genes and cognition are extremely indirect (as indeed they are between, for instance, sodium channels and cognition).

French mentions the FOXP2 gene, and this is an interesting example. A mutation of this gene produced an apparent specific deficit of speech and language with Mendelian inheritance in a British family. Computer simulations indicated that the human version of this gene may have appeared relatively recently on an evolutionary timescale, around the time language appeared in humans (see Marcus & Fisher Reference Marcus and Fisher2003). However, subsequent research indicated that behavioral deficits were more widespread than language (affected family members also showed deficits in oral-facial movements, aspects of the perception of rhythm, production of rhythmic movements of the hand, and on average lower IQ than unaffected family members) and a range of structural brain differences were observed when affected family members were compared to unaffected members or controls (including in sensorimotor cortex, inferior temporal cortex, inferior frontal cortex, supplementary motor areas, the planum temporale, caudate nucleus, insula, putamen, and cerebellum) (see Vargha-Khadem et al. Reference Vargha-Khadem, Gadian, Copp and Mishkin2005). There is undeniably an uneven cognitive profile in affected family members, yet the effects of the single gene mutation were far from specific on either brain or behavior.

Indeed, Plomin and colleagues have argued both that relations between genes and cognition are likely to be many-to-many, so that each gene influences many cognitive processes and each cognitive process is influenced by many genes (e.g., Kovacs & Plomin Reference Kovacs and Plomin2006; Plomin & Kovacs Reference Plomin and Kovacs2005; see also Rutter Reference Rutter2007), and that gene expression is likely to be distributed widely throughout the brain rather than in specific regions. Two of the most studied genes within cognition, COMT and BDNF, have basic neural functions related to neurotransmitter metabolism and neural growth respectively, and their expression is found across cortex, cerebellum, caudate nucleus, amygdala, thalamus, and corpus callosum in both humans and mice (Kovas & Plomin 2006). Given the widespread expression of genes in the cortex, it might seem reasonable to conclude that they exert broad sorts of constraints on the representations that emerge. Of course, this picture may change. The field of genetic neuroimaging, which studies where and when various genes are expressed in the brain, is advancing rapidly. Nevertheless, while we have no doubt that it will soon be commonplace to include genotype information in studies of cognition (where it will pick up some proportion of the variance in behavior, along with other experimental factors), we believe it premature to identify the exact constraints that genes place on current theories of cognitive development.

Two commentaries urge us to expand the neuroconstructivist framework in other directions, to incorporate evolution (Innocenti) and to cover lifespan changes in cognition rather than just development (Foster et al.). Innocenti argues that an evolutionary perspective prompts a greater emphasis on selectionism, where evolution selects certain neural and/or cognitive structures and development makes a further selection from possible cognitive architectures depending on the (proactively defined) environment in which the organism finds itself. In discussing the apparent opposition between selectionism and constructivism, Innocenti draws a contrast between mechanisms of evolution where “the power of selection is unchallenged” and the developmental time frame within which proactivity (or active exploration) is more important. Recent advances in evolutionary theory as applied to behavior make us less confident about this contrast. The fact that the activity of animals, guided by their embodied brains, changes their environment is well known. For an extreme example, by damming a river, families of beavers change their local habitat, flora, and fauna for several generations. However, this process of niche construction generates a feedback process in evolution that has only recently been fully recognized (Olding-Smee et al. Reference Olding-Smee, Laland and Feldman2003). The implications of this new perspective are that all aspects of niche construction, including cultural processes in humans, can influence the selection of genetic variation. Thus, the view is that there is no longer just passive selection of genes by the environment, but rather that species can be considered to create the environment that subsequently selects particular genes in future generations. We speculate that different individuals, or groups of individuals, from a species may create novel niches in a manner similar to proactivity dependence. Thus, it is possible that a similar dynamic balance between selectionism and constructivism occurs in evolution as occurs in individual ontogeny.

For us, one of the key questions in considering the role of evolution on human cognition is to understand the levers available to evolution to selectively change aspects of the brain and/or cognitive system. Work by Finlay and colleagues (e.g., Finlay & Brodsky Reference Finlay, Brodsky and Kaas2006; Finlay et al. Reference Finlay, Darlington and Nicastro2001) demonstrates that a large amount of structural variance in mammalian brains can be explained by a small number of factors producing nonlinear scaling across species. This is consistent with the availability of only a small number of fairly general genetic levers to change brain structure rather than a panoply of individual levers that control specific brain areas or specific cognitive abilities. In particular, a many-to-many mapping between genes and cognition would make fashioning a particular cognitive process via evolution problematic. Let us say a given process increased fitness so that its genes were more likely to propagate in the population. Under a many-to-many mapping, the favored genes also contribute to the development and maintenance of a large number of other cognitive processes. Selection for the one would therefore entail selection for the many. In short, for both evolutionary and genetic contributions to developmental theory, the key issue is the specificity of the influences on higher cognition.

We are in agreement with Foster et al. that the neuroconstructivist framework may be profitably extended to consider changes over lifespan, where we believe the principles we have identified would be equally relevant (see the work discussed in Section 1). Craik and Bialystok (Reference Craik and Bialystok2006) recently discussed the generalized lack of an interface between development and aging research, and we agree that it is desirable to make links between developmental and lifespan changes in cognition. We would hesitate to take a modular view of the development of more fundamental cognitive abilities but agree with Foster et al. that chronotopy may be an important characteristic of the changes observed in adolescence related to prefrontal cortical functioning. The control, coordination, and flexible use of cognitive abilities is contingent on the development of those abilities (see also sect. 2). We also find appealing the idea that aging may appear to be the converse of interactive specialization, with poorer performance marked by increases in the number of areas activated in the brain. Such parallels must be treated with caution, however, since there may also be differences: Aging has differential effects across brain regions that find no correspondence in development (see, e.g., Burke & Barnes Reference Burke and Barnes2006), and aging may alter different neurocomputational properties than those responsible for changes in network specialization during development (Li et al. Reference Li, Lindenberger and Sikstrom2001).

R6. Chronotopy, disorders, subjectivity, complexity, and formal methods

For reasons of space, we deal with a remaining set of questions more briefly. Issues revolving around chronotopy arose in several contexts. Bateson highlights the role of competition in effective reductions in plasticity, where experiences of one kind can exclude subsequent experiences of another kind (see Richardson & Thomas Reference Richardson and Thomas2008; Thomas & Johnson Reference Thomas and Johnson2006; Reference Thomas and Johnson2008, for further discussion of sensitive periods in development and their possible neurocomputational basis). As we saw above, Foster et al. discuss how the developmental changes in adolescence are reliant on the emergence of more fundamental cognitive abilities, and Shultz presents a possible computational formalism for exploiting previous knowledge in learning new abilities. Similarly, individual differences in aging indicate that the changes observed are to some extent dependent on the outcome of the developmental processes that preceded it (Stern Reference Stern2002). For us, the key point here is that chronotopy should not be a descriptive principle – simply identifying the order of developmental events – but should allude to the causal mechanisms that explain the necessity of the ordering.

On the topic of developmental disorders, Campos & Sotillo rightly identify that the neuroconstructivist perspective has implications for intervention. In particular, it is important to focus as closely on “scores within the normal range” or apparently normally developing abilities as on the impairments, both to verify that these abilities are genuinely developing normally and to evaluate their potential to serve as scaffolding for apparently weaker abilities. For example, individuals with Williams syndrome are known to have visuospatial difficulties that can affect their navigation skills. This can have an impact on their everyday life – for example, on making independent trips to the shops. Language is held to be a relative strength in the disorder. But is language in Williams syndrome sufficiently well developed to accommodate a verbal strategy for navigating to the shops and back? (See Thomas Reference Thomas2004 for a discussion of the relation of neuroconstructivist research to intervention).

Here is it worth noting that Ramus (Reference Ramus2004, p. 100) has characterized the neuroconstructivist approach as containing the following axiom: “no cognitive module in the adult mind can have originated in the newborn mind but must have been acquired through interaction with the environment,” with the logical corollary that in the case of developmental disorders, “a cognitive module cannot be congenitally disrupted.” We should stress that there is no such axiom. We are as keen as anyone to constrain our developmental accounts with as wide a range of empirical evidence from genetics, neurobiology, neuroimaging, computational modeling, and behavioral work as possible to build models that contain realistic constraints on the specificity of early deficits and their consequent effects on developmental trajectories. The neuroconstructivist perspective on developmental disorders is merely that a theory must specify starting conditions and the nature of the subsequent developmental process, and that both these aspects should be consistent with what is known from developmental cognitive neuroscience (Thomas & Karmiloff-Smith Reference Thomas and Karmiloff-Smith2002).

Innocenti points out that developmental disorders stemming from atypical environments can be equally informative about development. While this was a class of disorder we identified in Chapter 11 of Volume I, we did not pursue it in the original volumes. Innocenti is right to highlight the importance of evidence from, for example, studies of deprivation and impoverished environments. Animal work also indicates how enriched environments can influence neural development, for example for rats reared in environments of varying complexity (reviewed in Praag et al. Reference Praag, Schinder, Christie, Toni, Palmer and Gage2002; Rosenzweig Reference Rosenzweig1996). Enriched environments have been observed to produce changes in cortical thickness, with larger cell bodies, more glial cells, and higher dendritic and synaptic density, as well as an increased rate of neurogenesis. The role of impoverished and enriched environments in modifying (or failing to modify) developmental trajectories is a key avenue of investigation.

Lickliter suggests that we underplayed the importance of value and meaning in the formation and maintenance of cognitive representations – in essence, the subjectivity of the individual. We agree with this criticism in that subjectivity, and the value system that defines it, was left implicit in the notion of proactivity. The work of Sporns on robotics in Volume II emphasizes how important the concept of value may be in determining the effective environment for the individual (e.g., Sporns Reference Sporns, Mareschal, Sirois, Westermann and Johnson2007, p.184). Our review of behavioral genetics in the first chapter of Volume I noted that in twin studies, most of the variation in behavior attributable to the environment is unique to the individual (making identical twins look dissimilar) rather than shared (causing both identical and fraternal twins to look similar). Therefore, even when the environment appears similar (such as growing up in the same family or going to the same school), individuals experience it differently.

A number of the commentaries remark on our preference for the use of formal modeling to specify causal mechanisms underlying developmental change. In some cases, the comments are optimistic on the gains that formal modeling will provide (e.g., Bateson), in other cases, they are pessimistic (e.g., Lickliter). For Lickliter, the difficulty lies in closed or bidirectional causal loops between organism and environment at multiple levels, producing “mind-boggling combinatorial complexity” and “dizzyingly complex multidetermined phenomena.” In some ways, we agree with this view. The complexity of the ongoing interactions at multiple levels of description makes analysis extremely problematic. Pribram also highlights the implication of complexity in the context of nonlinear dynamics. Formal methods can at least offer a chance to understand the emergent properties of such interactions, for example as illustrated by Sirois and Mareschal's model of infant habituation (Reference Sirois and Mareschal2004). At the very least, formal models force the investigator to understand the complexity of the theory that is being proposed. A term like proactivity is easy to postulate at the qualitative level, but formal modeling uncovers serious issues associated with it. Allowing a learning system to determine the nature of its training set by its actions on an environment places it at great risk of instability. As Ferber recognizes, this has much in common with the problems faced by engineering control systems that operate via feedback (see, e.g., Mees Reference Mees1981). Without damping, a proactive system may oscillate between improving on different, inconsistent parts of the training set. Without time-sensitive changes of internal state, the system may lapse into locally stable solutions that are far from optimal for the problem it faces. The fact that formal approaches reveal such implications confirms that, however high the Everest of development may be, formal modeling leads us uphill.

R7. Omissions

As we stated in the preface to Volume I, too many great people have influenced our thinking for us to be able to acknowledge each directly. But we are grateful for the chance to acknowledge some of them here, in response to the influences recognized by the commentators. For example, French points out that we had not cited Hebb (Reference Hebb1949) – so pervasive was his influence on our thinking that we failed to cite him! French also notes the similarities of our machine-learning analysis of development (see Mareschal & Thomas Reference Mareschal and Thomas2007) to that proposed by Regier (Reference Regier1996), which we are happy to acknowledge. Similarly, Lickliter points out the importance of the work of Walter Freeman (Reference Freeman2001) on the relation between representation and meaning. While Bateson agrees with much of what we say, he chastises us for not properly acknowledging the contributions of many developmentally inclined ethologists and neuroethologists. Of course, it is a hazard in an interdisciplinary book such as this that experts from any single one of the constituent fields may feel that their discipline has not been adequately represented in the mix. However, this particular criticism stings hard because we consider ethology (as the behavioral development of animals in their natural environment) a critical and vital foundation stone for our book. For example, at the start of chapter 4 of Volume I on embodiment we clearly root this idea in classical ethology, giving examples from the work of Tinbergen, Ewert, and Beer among others. We fully acknowledge that there could have been many other names added to this list, but our primary objective was to then merge these ideas with those from robotics and artificial intelligence, and not to provide a scholarly review of the ethology literature. For reasons of readability, therefore, we gave only a few examples. However, we welcome the opportunity to clarify the seminal importance, and influence on our thinking, of ethologists such as Bateson, Oyama, Hinde, and Gottlieb (the latter two already being cited in the book). For at least one of the authors (M.H.J.), this is the single most important intellectual tradition that influences our work. That it continues to be so is evident from our discussion of niche construction in Section R4.

R8. Conclusion

We are glad to learn from the contributors how they believe the neuroconstructivist framework can be expanded. The Neuroconstructivism books are the first steps on a long road. We have aimed to provide a framework within which to think about development and the factors affecting it. We believe that this framework is a powerful one that integrates disparate fields and levels of description into a unified whole. Many researchers can contribute to this framework from their own areas of specialization. The commentators offer clear pointers of the directions in which neuroconstructivism can expand: a more detailed consideration of the theory of interactive specialization; specification of the relation between the emergence of partial representations and developmental changes in abstraction and cognitive flexibility, including the role of the prefrontal cortex; the important role of genetics in explanations of development; the function of value and subjective experience in proactivity; changes in cognition across lifespan, including the putative reorganization in adolescence and the processes underlying cognitive decline in the aging population; the effects of atypical environments in deflecting developmental trajectories; the implications of a neuroconstructivist analysis of developmental disorders for intervention; and, at the formal level, the implications of closed causal loops between the individual and his/her environment. With such avenues to explore, we believe that the multidisciplinary study of development has a rich future in the twenty-first century.