Are Dynamic Systems and Connectionist Approaches an Alternative to ―Good Old Fashioned
Lisa M. Oakes, Nora S. Newcombe, and Jodie M. Plumert
The question of whether dynamic systems and connectionist approaches represent a new, unified theory of cognitive development leads naturally to an evaluation of what these approaches offer beyond traditional approaches to cognitive development. In our view, dynamic systems and connectionism bring a renewed focus on emergence in thinking about developmental change. Although this notion is also central to traditional theories of Good Old Fashioned Cognitive Development (GOFCD) such as those of Piaget, Gibson, and Vygotsky, the idea was lost in the heyday of nativism. In this chapter, we trace the historical roots of concepts central to dynamic systems and connectionist approaches, make connections between GOFCD and dynamic systems and connectionism, and illustrate through examples how work from more ―traditional‖ perspectives has also revealed the process of emergence in developmental change. Finally, we consider connectionist and dynamic systems approaches to two classic areas of cognitive development, the A-not-B error and the balance scale problem, in the context of more traditional approaches to those problems. These comparisons allow us to examine how the new explanations refocus our attention on emergence and discuss the pitfalls that these approaches do and do not avoid.
The goal of this volume, and the conference that motivated it, is to determine whether connectionism and dynamic systems are two distinctly different theories of cognitive development or whether together they represent a paradigm shift in the field of cognitive development toward a single new grand theory. The three authors of this chapter represent relative outsiders to this discussion. We each have studied cognitive development from a blend of ―traditional‖ theoretical approaches, influenced by the theories of Piaget, Gibson, and Vygotsky, as well as by the broad range of theories that fall under the umbrella of information-processing approaches to cognitive development. We each represent a different blend of these theoretical frameworks, and we are all sympathetic to the general aims of and ideas behind connectionism and dynamic systems. We therefore have taken on the task of critically evaluating these ―new‖ approaches from the perspective of Good Old Fashioned Cognitive Development (GOFCD), with an eye toward understanding what connectionism and dynamic systems bring to the field as well as understanding the extent to which they differ from more ―traditional‖ approaches to cognitive development.
We have organized this chapter into four sections. In the first section, we discuss what connectionism and dynamic systems bring to the study of cognitive development. Because many of the chapters in the volume deal directly with this issue, this section is relatively brief. In the second section, we examine how connectionist and dynamic systems theories relate to other GOFCD theories of developmental change. In some sense, this section traces the historical roots of key ideas in connectionist and dynamic systems theories. In addition, we show through examples of systematic programs of work examining developmental change in cognitive processes how the ideas inherent in connectionism and dynamic systems are not unique, although, importantly, connectionist and dynamic systems approaches may make some of these ideas more explicit and central, in part because developing formal analytical methods for modeling change provides a new set of sharper tools to drive progress. Next, we evaluate the contribution of
connectionism and dynamic systems in more depth by examining explanations of two historically important issues in cognitive development: infants’ behavior in the A-not-B task and children’s
solutions to the balance scale problem. Here we evaluate how these new theories compare to more traditional explanations of children’s developing behavior in these tasks. Finally, we
consider how well connectionist and dynamic systems approaches address criticisms often leveled at other theories of cognitive development.
Dynamic systems and connectionist approaches to the study of change
The field of cognitive development is broadly concerned with how children’s thinking
evolves from the preverbal representations of an infant to the high-level conceptual abilities of a 16-year-old. The challenge, of course, is how to characterize and study such change and the causes of change. Dynamic systems and connectionist approaches to cognitive development explicitly focus on understanding change over time. Each is concerned with demonstrating through mathematical models and careful empirical studies how change occurs, not simply
documenting that change occurs. As a result, they bring a focus on mechanism to the forefront. Proponents of the two viewpoints are concerned with how systems self-organize, with organization arising from a less organized (or sometimes unorganized) state through real-time processes and the dynamic activity of the system. The two approaches conceive of this self-organization differently. For dynamic systems theories, developmental change is an emergent product of interactions among multiple components, occurring on many different timescales (Smith & Thelen, 2003). Theories adopting this framework emphasize multi-causality and self-organization emerging out of the real-time dynamics of the child’s own activity in a structured
environment (Smith & Thelen, 2003). For connectionist theories of development, reorganization emerges out of nonlinearities in learning (Marchman, 1997; Thelen & Bates, 2003), and new structures only emerge from the interaction of the existing structure and environmental input (Bates & Elman, 2002; Elman, 2005). From an outsider’s perspective, it is very difficult to
distinguish between these two ways of thinking about change; for both, self-organization and
emergent structure are a key feature of change, structural change emerges from activity that occurs in real-time, and developing systems exhibit high levels of variability during the process of change. Central to both connectionist and dynamic systems theories of development, therefore, is the explicit idea that new structures and behaviors are emergent products of multiple, interacting components. Moreover, in both styles of theorizing, change at longer time scales necessarily emerges from change at shorter time scales because all behavior is linked together through time. Perhaps most important, both approaches involve developing formal mathematical models of developmental change that provide a detailed level of specificity about those interactions (although it is important to point out that the two most influential volumes on connectionist and dynamic systems approaches to development did not include formal models; Elman, Bates, Johnson, Karmiloff-Smith, Parisi, & Plunkett, 1996; Thelen & Smith, 1994).
One consequence of these core ideas is that studies conducted within connectionist and dynamic systems frameworks often involve repeated observations of behavior over time—
although the time scale is often relatively short—within a single session or across sessions
separated by a few days or weeks. Because real-time change (i.e., the changes that happen from moment-to-moment) is intimately tied to change at longer time scales, development can be understood by observing change over many trials or epochs within a single experimental session, or over several sessions or epochs across several days or weeks (for examples see Samuelson, 2002; Spencer, Vereijken, Diedrich, & Thelen, 2000; Thelen, Corbetta, & Spencer, 1996). Because the work from a GOFCD perspective often does not have as an explicit goal uncovering the mechanisms of change, studies involving repeated observations over time are relatively rare (for a notable exception, see Siegler, 1996). Instead, GOFCD theorists generally use cross-sectional studies to document cognitive changes that occur over relatively long time scales (e.g., months or years). As others have noted, it is very difficult to examine mechanisms of change with cross-sectional studies. Thus, one contribution of the connectionist and dynamic system movements in cognitive development is to put the focus back on repeated observations over time.
It must be pointed out that this approach is not new to cognitive development—Heinz Werner
(1957) called it microgenesis, and Bob Siegler (1996) has strongly advocated and practiced this style of research over the past decade or more (see also Oakes & Plumert, 2002; Plumert & Nichols-Whitehead, 1996 for empirical examples of this approach). It should also be noted that despite their interest in the connections between changes on different time scales, most studies from a dynamic systems perspective examine change over relatively short time scales (such as trials or minutes). Thus, one weakness of many studies adopting a connectionist or dynamic systems framework is that they do not often examine changes over long time scales (Thelen’s
work on reaching would be an exception, e.g., Spencer et al., 2000; Thelen et al., 1996).
Another key contribution of dynamic systems and connectionist approaches to cognitive development are the tools they provide for studying the emergence of new structures or behaviors (Bates, Elman, Johnson, Karmiloff-Smith, Parisi, & Plunkett, 1998; Thelen, Schoner, Scheier, & Smith, 2001). As is evident from several of the chapters in this volume (see xxxxx, xxxxx, and xxxx), not only do connectionist and dynamic systems theorists value repeated observations over time, they also call for studies that seek to understand the processes that give rise to emergent behaviors. Such studies typically include perturbations or supports that change the organism-environment interaction. Usually, this entails manipulations of environmental structure (e.g., changing the salience of the B location and noting the effect on infants’ reaching for the A location in the A-not-B task, Diedrich, Highlands, Spahr, Thelen, & Smith, 2001), but sometimes it entails manipulation of organism characteristics (e.g., teaching children shape-based categories in the lab, presumably changing how they approach the word-learning context in general, and noting the influence of this change on their rate of vocabulary acquisition outside the lab, Samuelson, 2000). Observing how manipulations of either the task or the organism changes the resulting behavior leads to a better understanding of the processes that give rise to behavior. Again, this type of approach is not unique to dynamic systems and connectionism. There are many examples from GOFCD and other approaches (e.g., see Lickliter, this volume) that
explicitly engineer changes to either the organism or the environment to gain insight into developmental process (e.g., Oakes & Plumert, 2002; Plumert & Hund, 2001; Robinson, 2005).
As the chapters in this volume also make clear, these two theoretical paradigms have provided new mathematical or computational tools that make it easier to examine organism-environment interactions (see chapters xxxxxx, xxxxxx, xxxxx, this volume; and Bates et al., 1998). For example, the dynamic field theory allows researchers to directly test how hypothesized processes within the organism (e.g., memory or attention) and inputs from the environment (e.g., salience of perceptual information) interact to produce predictable patterns of behavior (e.g., Schutte, Spencer, & Schoner, 2003). Likewise, connectionist models of learning allow researchers to directly test how patterns of behavior emerge out of the interaction of simple processing units (e.g., Mareschal, Quinn, & French, 2002). Together, the conceptual and computational focus on how behavior emerges from interacting components offers a significant step forward in our understanding of developmental process.
In summary, a major contribution of these two approaches to the field of cognitive development is a focus on the mechanisms of change that lead to the emergence of new behaviors. Hence, any ―grand new theory‖ from either a dynamic systems or connectionist perspective
would seek to outline general principles that govern how new ways of thinking or behaving arise from multiple, interacting components. Importantly, this ―new‖ theory would make explicit this focus on the emergence of new behaviors through such interactions, and have as a central goal understanding those interactions rather than describing age changes in cognitive skills. In the next section we evaluate whether this would indeed be a new theory, and show how many GOFCD theories have also had this as a goal.
Is this a new way of thinking about developmental change?
GOFCD has long been interested in change, and much of the research and theorizing in the field of cognitive development is ultimately motivated by understanding developmental change. For example, information-processing theories, which have been criticized for focusing on
what develops rather than on mechanisms of developmental change (Thelen & Smith, 1994), have described change in terms of an increase in processing speed (Kail, 1986), the number of relations a child can keep in mind (Halford, Wilson, & Phillips, 1998), the capacity and duration of memory stores (Case, 1985), and the availability of strategies for solving problems (Siegler, 1996). Importantly, in each of these examples, mechanisms of changes are provided through increasingly thorough and detailed descriptions of what is developing.
Why then does the emphasis on change in dynamic systems and connectionist approaches seem to be so unique and novel? In the 1980’s and 1990’s, nativist theories dominated the study
of cognitive development. Such theories focused on identifying early emerging capabilities and not on how change occurs. This focus reflects, in part, the influence of Chomsky’s (1968) notions
that environmental events simply ―trigger‖ pre-existing behaviors (i.e., those specified in biology).
The highly publicized and influential work of people like Elizabeth Spelke (Spelke, Breinlinger, Macomber, & Jacobson, 1992; Spelke & Newport, 1998), Karen Wynn (1998), and Rochel Gelman (1978), was all aimed at showing high-level cognitive competence at an early age rather than documenting the mechanisms that produce changes in cognitive abilities. This led to the impression that the field of cognitive development had collectively lost an interest in
1understanding how change occurs.
However, understanding the mechanisms that lead to the emergence of new behaviors has a long history in the study of cognitive development and was central to the theories of Piaget, Gibson, and Vygotsky (a fact that is acknowledged by theorists from both connectionist and dynamic systems perspectives, see Bates & Elman, 2002; Thelen & Bates, 2003). Piaget, for example, proposed that new mental structures emerge through the dynamic interplay between the child’s developing cognitive structures and input from the environment. Similarly, for Gibson, changes in the organism lead to increased sensitivity to environmental structure, which in turn leads to changes in the organism at both neural and behavioral levels. Hence, change emerges out of cyclical organism-environment interactions over both shorter and longer time scales (Gibson,
1988; Gibson & Pick, 2000). For Vygotsky, new skills emerge at times when children are sensitive to social experiences that allow them to try out new ways of thinking and acting, sometimes referred to as the zone of proximal development (Wertsch, 1985). The notion that adult guidance must be developmentally appropriate necessarily implies that cognitive change emerges out of the interaction of the child and the social environment. Thus, the idea that behaviors emerge through interactions between the organism and the environment has been central to our understanding of development for quite some time, and these historical views on emergence have played an important role in the application of dynamic systems and connectionist frameworks to understanding development.
Importantly, however, the ideas about emergence in the theories of Piaget, Gibson, and Vygotsky have largely been lost or ignored over time, even by theorists who came from these traditions. Modern theorists whose work originated in a Piagetian tradition shifted focus from understanding how cognitive structures emerge out of the interaction of the child and the world to an almost exclusive focus on the cognitive structures or concepts themselves (Flavell, 1970). Gibsonian theorists have shifted away from viewing affordances as an emergent property of the interaction between the organism and environment to viewing affordances as objective properties of the environment - "information available about surfaces, places, obstacles, and things as well as about oneself" (Gibson, 2003, p. 293). Likewise, socio-cultural approaches to cognitive development have almost exclusively focused on the social environment and have had very little to say about how organism characteristics interact with social structure to produce changes in thinking. We believe these shifts in theoretical perspectives over time have occurred because the notions of interaction and emergence are very difficult concepts. It is difficult to think about
behavior or thinking as being simultaneously determined by organism characteristics and environmental structure. It is much easier to assign causal priority to one or the other, rather than to both at the same time.
An additional issue is that arguments about the emergence of new ways of thinking or
behaving from a GOFCD perspective have been made primarily at a conceptual level. Although this theorizing has led to advances in our understanding of the mechanisms of development, the hypothesized mechanisms have often been difficult to test. One significant contribution of connectionist and dynamic systems theories of cognitive development is to illustrate how interaction and emergence can be simulated with sophisticated formal models. Indeed, one of the most exciting products from this movement is the collection of mathematical models that simulate the emergence of qualitatively different stages of behavior from multiple, interacting components. Although computational models can oversimplify the complexity of cognitive processes, they offer an important step forward in formalizing and testing ideas about cognitive change.
Because interaction and emergence are difficult to conceptualize, many studies of cognitive development from a GOFCD perspective may appear to be ―merely‖ descriptive at first blush, rather than revealing processes of cognitive change. However, the rich descriptions of cognitive development amassed over the last several decades have actually provided considerable understanding of the mechanisms of change, although the explicit goals of these investigations may not have always been to uncover such mechanisms. Thus, we have made more progress in our understanding of mechanisms of change than we often credit ourselves for, and have even advanced our understanding of how behavior and structure emerge from organism-environment interactions. This progress has not occurred despite a focus on description, but rather is intimately tied to the descriptive success. In part, the two enterprises are linked because the rich description of change strongly constrains the search for explanation. But even more than that, rich description focuses attention on plausible mechanisms and explanatory principles. These mechanisms can then be tested in studies that may themselves seem (on the surface) descriptive. That is, investigators often test explanatory ideas by evaluating predictions regarding ―descriptive‖ questions, such as how manipulations in task characteristics or in environmental input will affect children’s behavior or cognitive abilities.
Consider an example of a cognitive change: the emergence of configural, as opposed to featural, processing of visual stimuli. We may simply ask: at what age do people process upright and inverted faces differently (an effect that has been interpreted as a hallmark of configural rather than merely featural processing)? This may seem a supremely descriptive question (although an improvement over a global question such as when children can process faces). Yet, the answer revealed by research asking this descriptive question turns out to be quite complex. On the basis of classic work by Carey and Diamond (1977), we would conclude that this change occurs at around 10 years of age, and that that what propels a change to configural processing are prolonged developmental changes—maturational and experiential factors that occur over many
years. Evidence that configural processing is in place by 4 or 5 months (Bhatt, Bertin, Hayden, & Reed, 2005; Turati, Sangrigoli, Ruel, & de Schonen, 2004), in contrast, led to a very different account of what factors influence a change from featural to configural processing. If configural face processing emerges early, then infants’ experience with faces during the period during which the visual system is organized becomes a potentially important factor contributing to that change. Indeed, early restriction of visual experience, as occurs when infants are born with congenital cataracts, has an enduring influence on face processing (Le Grand, Mondloch, Maurer, & Brent, 2004)—an observation that might appear to be merely a descriptive fact, but that is also a test of a mechanism.
Even further specificity turns out to be possible. It turns out that there are three types of configural processing, and they mature at different rates (Mauer, Le Grand, & Mondloch, 2002). Although some kinds of configural face processing may be present in infancy, others may not reach adult levels until 14 years. In addition, effects of early visual deprivation can co-occur with surprising levels of adult plasticity (Maurer, Lewis, & Mondloch, 2005), as may arguably be seen in the advent of configural sensitivity to non-face stimuli (Gauthier & Tarr, 2002). Taken individually, each of these investigations could be characterized as descriptive—they illustrate
how the behavior unfolds over time. Taken together, however, the data (at a minimum) strongly