7.7: Internal Representations
- Page ID
One of the key properties of classical cognitive science is its emphasis on sensethink-act processing. Classical cognitive scientists view the purpose of cognition as planning action on the basis of input information. This planning typically involves the creation and manipulation of internal models of the external world. Is the classical sandwich (Hurley, 2001) a mark of the classical?
Sense-think-act processing does not distinguish classical models from connectionist networks. The distributed representations within most modern networks mediate all relationships between input units (sensing) and output units (responding). This results in what has been described as the connectionist sandwich (Calvo & Gomila, 2008). Sense-think-act processing is a mark of both the classical and the connectionist.
While sense-think-act processing does not distinguish classical cognitive science from connectionism, it may very well differentiate it from embodied cognitive science. Embodied cognitive scientists have argued in favor of sense-act processing that abandons using internal models of the world (Pfeifer & Scheier, 1999). The purpose of cognition might not be to plan, but instead to control action on the world (Clark, 1997). Behaviour-based robots arose as an anti-representational reaction to classical research in artificial intelligence (Brooks, 1991). The direct link between perception and action—a link often described as circumventing internal representation—that characterized the ecological approach to perception (Gibson, 1979; Turvey et al., 1981) has been a cornerstone of embodied theory (Chemero, 2009; Chemero & Turvey, 2007; Neisser, 1976; Noë, 2004; Winograd & Flores, 1987a).
The distinction between sense-think-act processing and sense-act processing is a putative differentiator between classical and embodied approaches. However, it is neither a necessary nor sufficient one. This is because in both classical and embodied approaches, mixtures of both types of processing can readily be found.
For example, it was earlier shown that the stored program computer—a digital computer explicitly designed to manipulate internal representations—emerged from technical convenience, and did not arise because classical information processing demanded internal representations. Prototypical classical machines, such as the Turing machine, can easily be described as pure sense-act processors (Wells, 1996). Also, earlier electromechanical computers often used external memories to scaffold processing because of the slow speed of their componentry.
Furthermore, prototypical classical architectures in cognitive science appeal to processes that are central to the embodied approach. For example, modern production systems have been extended to include sensing and acting, and have used these extensions to model (or impose) constraints on behaviour, such as our inability to use one hand to do two tasks at the same time (Kieras & Meyer, 1997; Meyer et al., 2001; Meyer & Kieras, 1997a, 1997b, 1999; Meyer et al., 1995). A production system for solving the Towers of Hanoi problem also has been formulated that uses the external towers and discs as the external representation of the problem (Vera & Simon, 1993). Some have argued that the classical emphasis on internal thinking, at the expense of external sense-acting, simply reflects the historical development of the classical approach and does not reflect its intrinsic nature (Newell, 1990).
Approaching this issue from the opposite direction, many embodied cognitive scientists are open to the possibility that the representational stance of classical cognitive science may be required to provide accounts of some cognitive phenomena. For instance, Winograd and Flores (1987a) made strong arguments for embodied accounts of cognition. They provided detailed arguments of how classical views of cognition are dependent upon the disembodied view of the mind that has descended from Descartes. They noted that “detached contemplation can be illuminating, but it also obscures the phenomena themselves by isolating and categorizing them” (pp. 32–33). However, in making this kind of observation, they admitted the existence of a kind of reasoning called detached contemplation. Their approach offers an alternative to representational theories, but does not necessarily completely abandon the possibility of internal representations.
Similarly, classical cognitive scientists who appeal exclusively to internal representations and embodied cognitive scientists who completely deny internal representations might be staking out extreme and radical positions to highlight the differences between their approaches (Norman, 1993). Some embodied cognitive scientists have argued against this radical polarization of cognitive science, such as Clark (1997):
Such radicalism, I believe, is both unwarranted and somewhat counterproductive. It invites competition where progress demands cooperation. In most cases, at least, the emerging emphasis on the roles of body and world can be seen as complementary to the search for computational and representational understandings. (Clark, 1997, p. 149)
Clark (1997) adopted this position because he realized that representations may be critical to cognition, provided that appeals to representation do not exclude appeals to other critical, embodied elements: “We should not be too quick to reject the more traditional explanatory apparatuses of computation and representation. Minds may be essentially embodied and embedded and still depend crucially on brains which compute and represent” (p. 143).
The reason that an embodied cognitive scientist such as Clark may be reluctant to eliminate representations completely is because one can easily consider situations in which internal representations perform an essential function. Clark (1997) suggested that some problems might be representation hungry, in the sense that the very nature of these problems requires their solutions to employ internal representations. A problem might be representation hungry because it involves features that are not reliably present in the environment, as in reasoning about absent states, or in counterfactual reasoning. A problem might also be representation hungry if it involves reasoning about classes of objects that are extremely abstract, because there is a wide variety of different physical realizations of class instances (for instance, reasoning about “computers”!).
The existence of representation-hungry problems leaves Clark (1997) open to representational theories in cognitive science, but these theories must be placed in the context of body and world. Clark didn’t want to throw either the representational or embodied babies out with the bathwater (Hayes, Ford, & Agnew, 1994). Instead, he viewed a co-operative system in which internal representations can be used when needed, but the body and the world can also be used to reduce internal cognitive demands by exploiting external scaffolds. “We will not discover the right computational and representational stories unless we give due weight to the role of body and local environment—a role that includes both problem definition and, on occasion, problem solution” (Clark, 1997, p. 154).
It would seem, then, that internal representations are not a mark of the classical, and some cognitive scientists are open to the possibility of hybrid accounts of cognition. That is, classical researchers are extending their representational theories by paying more attention to actions on the world, while embodied researchers are open to preserving at least some internal representations in their theories. An example hybrid theory that appeals to representations, networks, and actions (Pylyshyn, 2003c, 2007) is presented in detail in Chapter 8.