A. Riegler, M. Peschl & A. von Stein (eds.)
Understanding Representation in the Cognitive Sciences
Plenum Academic / Kluwer Publishers
Overview of Contributions
The first papers address the epistemological status of representation and possible categorizations. These rather general considerations about representation are followed by contributions that focus on computational approaches, especially on how meaning can be simulated in artificial devices, and how meaning can be grounded. Solutions are offered on various levels, among them symbol grounding and system-theoretical considerations to representation. Various disciplines independently arrive at emphasizing the importance of actions for representations and the necessity to close the sensory-motor loop; examples from the perspective of computational approaches ("embodiment" into an environment), the neuroscientific perspective ("top-down" processes), the psychological perspective and others are presented. Several papers address the role of social interactions--in particular language--as a means of stabilizing systems, and in the emergence of meaning. The deep epistemological implications connected with the above considerations are discussed in the final chapter of the volume dedicated to constructivist approaches. Following the idea of a true interdisciplinary approach, the contributions to the various sections are organized exclusively with respect to thematic content rather than to scientific discipline. The remainder of this overview presents the papers and their mutual relationship in more detail.
In the first chapter, dedicated to theoretical considerations, Georg Dorffner provides an introduction to the problem of representation by defining three types of representation. He shows that connectionism solves several problems of representation and actually helps to abandon one of these types of representation. Similarily, Alfredo Pereira investigates two types of the representations used in the neurosciences, perceptual and executive processes. Anthony Chemero tries to classify possible critiques regarding representations. He presents two types of anti-representationalism: either the representing structure can be questioned ("empirical anti-representationalism"), or the represented structure can be questioned ("metaphysical anti-representationalism"). Evidence for the first comes e.g. from connectionism which has demonstrated that brain-like structures can function without explicit representations. Arguments for the second are of a philosophic nature denying the possibility of speaking of fixed entities of the world, and, instead, proposing that the "things" are defined only with respect to their user and thus cannot be "mapped" onto the brain of that user. Matthias Scheutz investigates the ontological status of representations, and questions whether they are entities on their own; he argues that whether something counts as representation of something else is dependent on the level of description. By talking about a mental structure representing something in the world, one implies that these two things exist on two different ontological levels. This is not true, however, since the mind and the things of the world are both entities within our cognitive experience. Thus, the problem of representation arises only as artifact of our describing the phenomenon of cognition. Daniel Hutto explores the difference between non-conceptual representations and conceptual representations and asks the question whether it makes sense to define representation on the lowest level of abstraction or on the highest level.
Is it possible to simulate meaning or cognition? The issue of computation to representation is presented in the chapter "Computational Approaches". The paper by Georg Schwarz draws conclusions for the functioning of the brain from a computational perspective. From a philosophical point of view, William Robinson tries to address a controversy in computational approaches to cognition by introducing a distinction between cognition and cognitive abilities. Robert French makes the argument that creating representations cannot be separated from manipulating them. Moreover, he emphasizes the context-dependency of linguistic terms (one of the reasons that computational linguistics is non-trivial matter) and, as a consequence, the distractive character of context-laden representations. Andrew Coward proposes that natural pressures have resulted in biological brains having simple functional architectures. The type of architecture constrains the type of representations which are possible.
While connectionism overcame some of the problems derived from considering the brain as a von Neuman Computer, serious new questions came into focus. With simulating representations in neuronal networks it became even more obvious that the problem to be solved was that the meaning was put into the system by the engineer. Whether the state of the network would stand for something is defined by an external observer but is not inherent to the system. This problem of the foundation of meaning is referred to as the grounding problem, which is treated in the chapter on "See Symbol Grounding and Language". Tom Ziemke provides a review about the grounding problem and it's two proposed solutions. The first solution of grounding is grounding a symbol to an input, which was first proposed by Stevan Harnad. The papers by Nathan Chandler et al. and Christian Balkenius & Simon Winter present such models on grounded language systems. As described in the paper by Ziemke, this, however, is only a partial solution; the real breakthrough is only obtained when a robot is really embedded into an environment and interacting with it. Examples of this kind of grounding are given by Mark Wexler and Ralf Möller (see below).
Not only in classical AI but also in connectionism, meaning and representations are mostly of a static nature. Thus, although distributed, representations are still very much treated like "boxes". Cognition, however, is dynamic and it needs time (see also discussion, The Relevance of Time for Representations). Systems theory provides a framework that encompasses these issues (chapter "Cognition as a Dynamical System"). From a psychological point of view, Pim Haselager provides a review about the relevance of dynamical systems theory to the topic of representation and its (philosophical) classification in the context of behaviorism. Marius Usher & Ernst Niebur present neuroscientific evidence and a model demonstrating that neural representations are active processes which can mediate contextual computation and bind relational properties. Ken Mogi discusses the concept of time and causality in the relation of perception and neural firing. His principle of interaction simultaneity provides an explanation for the origin of subjective time. Steve Bressler presents a complex-systems theory approach to representation and cognition. Inspired by empirical findings of local and large scale cortical interactions, he derives a dynamical view of cortical representation based on interareal pattern constraints and large-scale relaxation.
In a related realm, away from representations understood as projections of environmental entities onto mental structures, several disciplines independently discovered the importance of actions for representations. They propose to close the sensory-motor loop. (For this topic of parallelisms in scientific/philosophical developments: see the comment by Wolf Singer in the discussions, Feedback Between Concepts and Theory-Development). In modeling, the grounding problem described above introduced the relevance of interaction for representations; embodied systems that interact with their environment in a sensory-motor loop are created. Philosophically, constructivism and the theory of autopoietic systems demonstrated that what we perceive as representations of a predefined external world is better described as constructions that serve to successfully generate behavior. In neuroscience, the old concept that environmental objects are mapped onto neurons in a feed-forward manner is challenged by anatomical and physiological evidence. It has been shown that information does not exclusively flow from the sensors to the higher brain areas, where the representations are then thought to be used for "thinking" and the final activation of motor programs ("bottom up"). Rather, massive activity is transported the other way around ("top-down"). Thus motor activity, i.e., planned actions, can be viewed as supervising the incoming signals and thus structuring cortical representations themselves. Psychophysical evidence, such as the figure-vase ambiguity, show that perception is ambiguous, dependent on priming and expectancy. Thus, several approaches come to the conclusion that perception is not a passive, feedforward process of mapping but an active construction, where the need for action creates an hypothesis about the environment which is then compared to the incoming sensory signals. The chapter "Relevance of Action for Representation" is dedicated to these problems.
The idea of "the inversion of sensory processing by feedback pathways" was introduced by Erich Harth . In his paper, he describes neurophysiological evidence for a scenario where processing is not from the periphery to the higher brain areas but rather the other way around. He presents a model where a cortical hypothesis ("internal sketchpad") might be controlled by the incoming signals in a hill-climbing algorithm. From the perspective of neurobiology, Astrid von Stein interprets the cortical architecture with its feed-forward and feedback connections as a medium of interwoven sensory-motor loops interacting on different levels of a hierarchy. Representations do not "sit" on either end of the connections, but are stable states in these sensory-motor loops. This is consistent with Piaget's finding that an object is defined through the perceived change in sensation that an action performed on the object induces. According to this approach, what we perceive as entities is neither in the world nor in the brain but in the interaction between both. Based on the idea of "perception through anticipation", Ralf Möller performs a simulation with a similar approach: he presents a system where the representation is not a building block within the system, with perception being a projection from outside, but where perception is a process of anticipating the sensory consequences of actions. Tom Wexler gives empirical evidence for the importance of actions in representations using both psychophysical examples and simulation (representation as strategy for acting rather than as a mapping of the environment). He demonstrates examples where previously unresolved problems (e.g. the "n-parity problem") can be solved if the learning system is directly interacting with the environment in a sensory-motor relation, i.e., if the environment is included into the representation.
When a first step in defining representations on a new ground is to consider the sensory-motor coupling with an organisms's environment, a second step is to consider the coupling not only with a static environment but with other representing systems. The role of social interactions--in particular language--for stabilities within a socio-cultural network, and the emergence of representations and "meaning" is a subject which is further treated by several authors in the chapter "Communication and Social Coupling". From a social-psychological point of view, Richard Eiser asks the question how individuals acquire knowledge or representation of social reality that adapt with new experience, and also how groups can coordinate their behavior in a manner adaptive to their environment. He presents a theory about the self, where the self is seen as the product of the cognitive system at work, i.e., as explanatory model for already performed or to-be-performed behavior. Tom Routen demonstrates that the advantage of the adaptive behavior approach over classical AI is only a partial one, and that language has to be included. He introduces the ideas of Bourdieu, and McDowell, who take language not as a static set of words and grammatical rules but as a "way of life", a Habitus. Habitus (Bourdieu) is "a system of dispositions which serves to set the individual's understanding and actions in agreement with its social and physical environment". Since Habitus is installed by socialization, it will be the same within a group, i.e., the experiencing will be similar and a common world-view will develop. Thus, language as a Habitat has to be included into models of cognitive agents. A prototypical example of representation, language and semantics is given by Peter Gärdenfors . In contrast to realist theories of semantics he suggests that it is not reasonable to speak about the relation between signs and things in the outer world. Instead, he argues that it makes sense to speak about the relation between signs (i.e., "linguistic expressions") and concepts, i.e., meaning in the head of the individuals. That linguistic expressions mean the same for everybody is not because they map a fixed world but because they emerge during interaction from the concepts in the heads of the language users. Examples of neurophysiological correlates of language processing are given by Sabine Weiss et al. The problem of transferring meaning between two languages--the problem of translation--is considered by Hanna Risku.
Till now we referred to representation as mental-environmental stabilities that have evolved to successfully guide the behavior of organisms; objects would in this sense be considered as adaptive constructs in this organism-environment relation. This, however, does not address the question of the subjective experience of representations, i.e., the awareness that there are things in the outside world, or the awareness of ones self as active part in the interaction with the world. The chapter "Qualitative Aspects of Representation and Consciousness" deals with this problem of consciousness. Wolf Singer addresses the issue of conscious experience from a neuroscientific perspective, ending with the conclusion that properties such as self awareness and consciousness cannot be explained in terms of neuronal correlates alone, but only by additionally considering the (social) interactions among brains. His position is, thus, both dualistic and materialistic, saying that consciousness is housed by a neuronal substrate but can never be explained by neuroscience itself since it considers a phenomenon which cannot be found in single brains (See also discussions, Consciousness and Language as Second Order Phenomena). Michael Pauen stresses the importance of structural similarities in representations: differences in the world have to scale to differences in the neuronal activities. Qualia is proposed to serve as the aspect organizing these structures. The problem of qualia dissolves if one abandons the artificial distinction between subjective experience and objective outside world.
Finally, the chapter "Constructivism" concludes with the deep epistemological implications the above considerations present. If we agree that what is represented in the brain is not just a projection from entities in the outside world but rather a repertoire of hypothesis about future actions, grouped and structured in an efficient way to successfully act--what, then, is that reality out there?
Ernst von Glasersfeld has a radical approach to that question: he argues that what we perceive as outside world is nothing but constructions of our cognitive apparatus. In his article he proposes to not use the word re-presentation at all, since it is commonly understood as a "copy" or "reproduction" suggesting the reference to an original. While this implies that we can compare a mental experience to something outside our mental experience--the outside world--he argues that this is not possible since mental content cannot refer to anything but to further mental content. He suggests to use the word "presentation" instead, to imply that it is a proposal the brain offers to the outside world in adapting to that world; a world, however, that in principal can not be known. Empirical evidence for the constructivist nature of representation is presented by Sverre Sjölander . He investigates the subject from an evolutionary perspective and describes how object-centered (invariant) representations develop from simple input-output processing in lower animals. He shows that the evolutionary purpose of sensory processing is not a mapping of the world but producing adaptive behavior. Annika Wallin shows that the fact that we have two types of representations--immediate perceptions and re-presentations such as during imagery--creates a serious problem for constructivism: according to constructivist theories, the concepts constructed to sub-serve adaptive behavior are adjusted according to the constraints presented to the system.
As this overview shows, our goal was to bring together interdisciplinarily working scientists in order to contribute to this volume. We believe that our selection of papers will help foster a better understanding and fruitful application of the ubiquitous nature of representation.