Loading...
Please wait, while we are loading the content...
Similar Documents
Something old, Something new: Extending the classical view of representation
| Content Provider | Semantic Scholar |
|---|---|
| Author | Markman, Arthur B. Dietrich, Eric |
| Copyright Year | 2000 |
| Abstract | ly is actually much easier when embedded in an actual situation. The goal of cognitive science, in this view, is to understand how agents structure their environment in order to solve complex tasks. 3. Embodied cognition Related to the situated action approach is embodied cognition. On this view, not only is it crucial to think about the contexts in which cognitive processing occurs, it is also necessary to build agents that actually interact in real environments [32, 33, 34, 35, 36]. Building real agents suggests ways that the environment can be exploited to solve difficult problems. Furthermore, while there may be many possible ways of representing information, the space of potential representations may be much narrower when the agent must achieve sensorimotor coordination. Thus, this view explicitly rejects the idea that cognitive theories can ignore perceptual and motor systems. There are many ways that the environment can be exploited to solve difficult problems. From Gibson's classic work on perception forward, scientists have demonstrated that the visual system is sensitive to information in the environment that provides information relevant to an organism's goals. For example, many species are able to use optic flow to gauge their direction and speed of motion. In addition, the vestibular systems provide information about linear and angular acceleration that can be used to augment visual information in the construction of cognitive maps [37, 38, 39]. Sometimes building a real agent can also lead to simple solutions to potentially difficult problems. For example, Pfeifer and Scheier [36] describe a robot that is able to Representation in Cognitive Science 13 distinguish large cylinders from small ones. This classification is accomplished by providing the robot with simple motor routines that allow it to follow walls and therefore to circle around objects. When the robot circles a small cylinder, the ratio of the speed of the outside wheel to the inside wheel is higher than when it circles a large cylinder. By using sensors that provide information about the speed of its wheels, the robot is able to perform a classification task without an elaborate visual system. Finally, Glenberg (1997) suggests that understanding memory requires attending to the function of memory within an organism. Many forms of memory require little effort, such as the perceptual priming observed following the presentation of a stimulus or the ability to point to the location of an object in space when the organism is navigating through that space. Glenberg [35] argues that these forms of memory are what permit organisms to carry out actions in the world. More effortful forms of memory require suppression of current input, which is what makes them more difficult to use [40]. Finally, he argues that language comprehension involves representing information as if the comprehender were going to act in the situation. In each of these cases, it is assumed that understanding cognition requires focusing on the relationship between an embodied organism and its environment. The embodied cognition approach has had great success at building very simple machines that navigate through environments and avoid obstacles. These agents are even able to perform simple tasks like picking up cans or classifying simple objects [34, 36]. Representation in Cognitive Science 14 The claim these researchers make is that all of cognition, including higher cognition, can be successfully modeled using this bottom-up approach. 4. Dynamical systems A final challenge to traditional assumptions about representation has come from proponents of dynamical systems as explanations of behavior [41, 42, 43]. Dynamical systems are systems of nonlinear differential equations that can be used to describe aspects of behavior (see Norton [44] for an introduction). On this view, a central problem with traditional approaches to representation is that they have discrete and enduring components. Dynamical systems do not involve discrete symbols. In a dynamical system, there is a current state consisting of the values of some set of control variables. There is also a set of equations that combine the control variables to govern how the system changes over time. Thus, the two key aspects of dynamical systems are that they involve continuous change in the values of the control variables, and that this change occurs continuously in time. Hence, dynamical systems assume that representations are time-locked to information in the represented world. As the state of the represented world changes, the representation changes as well. As an example, Kelso [41] describes a number of studies involving the coordination among limbs. For example, put your hands in a fist and place them in front of you. Then, extend the index fingers on both hands. Now flex and extend these index fingers in synchrony, increasing the speed of movement. Most people are able to maintain this coordination, even at high speeds. In contrast, try this same task, except Representation in Cognitive Science 15 flex one index finger as you extend the other. As people increase the speed of this movement, it often becomes difficult to maintain, until finally, they end up performing the first movement (flexing and extending both fingers at the same time). Kelso is able to describe this movement, as well as many more complex kinds of motor coordination using dynamical systems. Further, he makes a convincing case that this type of explanation is superior to an explanation of these behaviors involving other types of representations. In this model, the state of the system changes through time as the positions of the fingers change. Thus, this model contains no enduring representations. Some researchers have argued that this success in describing motor behavior can be extended to all of cognitive processing [42, 43]. They suggest that dynamical systems have two advantages over other approaches to cognition. First, by focusing on processes that evolve continuously, they are able to account for the plasticity of cognition. Second, it is assumed that continuous processes allow dynamical systems to account for the fine details of processing, which in turn allows them to account for individual differences. This focus on individual differences contrasts with much research in cognitive science, which focuses on commonalities in behavior across individuals. Semantics and representation The four alternative approaches to representation have focused primarily on lowlevel perceptual and motor processes. They have not had success at explaining higherlevel cognition. There is a good reason for the problems these models have with complex cognitive processes. To some degree, each of the alternative approaches ties representations to perceptual and motor pathways. On the positive side, this coupling of Representation in Cognitive Science 16 representation with perceptual and effector systems provides a basis for the semantics of the representation. In particular, one important way that representations come to have meaning is for them to correspond to something external to the agent. On the negative side, using correspondence as the primary basis for semantics is more likely to be successful for perceptual and motor processes than for high-level cognition. People's ability to represent abstract concepts involves a second aspect of semantics: functional role. That is, the meaning of a representational element is also determined by its relationship to other representational elements. If a theory of representation focuses primarily on correspondence, then processes that require functional role information will be difficult to explain. So, how should the classical view be extended? The classical view of representation has served as the basis for research in cognitive science since the late 1950s. This is a long time for a single framework to hold sway in a young science. Nonetheless, there is no reason to abandon the classical view yet. None of the problems identified by advocates of the four alternative approaches are fatal to the classical approach to representation. Instead, they are simply signs of growing pains. All of the approaches to representation discussed here agree on the fundamental assumption that cognitive processing involves internal mediating states that carry information. Thus, the exploration of representation can be fruitfully described as an Representation in Cognitive Science 17 examination of the types of properties that must be added to the basic concept of a mediating state in order to capture cognitive processing. Each of the alternative approaches essentially highlights particular properties that must be added to mediating states in order to account for cognitive processing [5]. Thus, the remaining assumptions of the classical view all require some change in light of the issues raised by alternative approaches, but it is always a change in scope. Not all representations are enduring, not all are symbols, not all are amodal, and not all are independent of the sensory and effector systems of the agent. The assumption that some representations are amodal is the one that requires the most future scrutiny. The studies described in the section on perceptual symbol systems suggest that tying representations to specific modalities may provide the basis for considerable flexibility in cognitive processing, and may even account for the use of abstract concepts. While it is too early to argue that cognitive science can dispense with amodal representations, it may be able to go a long way without them. The other three assumptions of the classical view are likely to survive intact for most aspects of higher cognitive processing. The assumption that cognitive systems have enduring states was challenged both by situated action and dynamical systems approaches. The situated action approach captures the important insight that many aspects of th |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://bingweb.binghamton.edu/~dietrich/Papers/ComputationalismRep/ExtendgClassicVuOfRep.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
