Markman & Dietrich (1998) argue that the notion of a "representation" should be replaced by the less contentious one of a "mediating state." They also believe that many mediating states will have the property of being inherently rule-governed. In arguing for this conclusion Markman & Dietrich ignore their own observation that being described by rules is not the same as using them inherently.
"Many systems can be described by rules, but that is not the same thing as using rules to carry out a process." (Markman & Dietrich, 1998, paragraph 61)
1. Markman & Dietrich (1998) suggest replacing the notion of representation with that of a mediating state. They believe all cognitive scientists will agree that mediating states are necessary both to explain cognition and to distinguish cognitive science from behaviorism. There is much to welcome in Markman & Dietrich's analysis of mediating states, though the notion itself is not new. Guthrie (1935), Osgood (1953), and Mowrer (1960), among others, developed the idea of mediating states in an S-R framework, and Osgood in particular applied his ideas in psycholinguistics. Indeed, such audacity led Jerry Fodor (1965) to question whether meaning could be a mediating state. Furthermore, connectionism, which many see as central to current cognitive science, is a better developed version of the associationism that formed the basis of behaviorism. Connectionism allows mediating states, but are these complex enough to support behaviors such as language use and complex problem solving?
2. My central concern here is with another aspect of mediating states: whether they are rule-governed (Markman & Dietrich, 1998, section C.5). Behavior must be rule-governed if cognitive science is possible. If there were not some order in the world, it is difficult to imagine how we could be here to talk about it. Furthermore, given the truth of Turing's thesis, "it MUST be possible in principle to model cognition using rules" (Markman & Dietrich, paragraph 63). But to model the behavior of a cognitive system is not necessarily to provide an account of its internal structure.
3. There is, however, a view, with antecedents in classical AI and the work of Chomsky and Piaget, "that behavior is inherently rule-governed" (Markman & Dietrich, paragraph 55). The crucial term here is "inherently". Chomsky's view, for example, is that the rules the system follows are part of its make-up (e.g., Chomsky 1980). This notion of rule-governed behavior, therefore, conflates the idea that the behavior can be described systematically with the idea that the rules that generate the behavior, in an abstract sense, also function concretely as part of the mechanism that produces the behavior.
4. As Markman & Dietrich note, a system that can be described as obeying rules need not make use of those rules in generating its behavior. In classical physics, there is little temptation to see objects (e.g., thermostats) as having internal structure that corresponds to rules they follow. Why, then, has such a claim been made in the cognitive sciences? Why, indeed, has it so often been taken as an obvious truth? One difference between classical physics and the cognitive sciences is that the sets of rules (or laws) that its objects of study follow (or obey) are relatively simple (e.g., Newton's laws of motion). Furthermore, objects that follow those laws are diverse. It is hard to imagine that they all have internal structure that corresponds to those laws. By contrast, many of the cognitive functions studied in classical AI and linguistics are entirely, or almost entirely, exclusive to humans (natural languages, chess playing, etc.). And it is assumed that the same or similar brain mechanisms underlie those behaviors in different individuals.
5. Not only are the sets of rules that people follow in performing cognitive tasks themselves complex: the outputs of those rules are likewise complex. Their structure has to be described, as does the fact that people can make judgments about those structures. One corollary of this complexity is that the rules describing, say, human languages are difficult to formulate, and they are particularly difficult to formulate as generative mechanisms, in the sense introduced by Chomsky. Computer programs can help, and it is tempting to see the computer plus its program as a (probably imperfect) copy of the human language processing mechanism.
6. A further complication is that we have to explain not only what structures sentences have, but also how people understand and produce those sentences in spoken or written utterances. Traditional AI systems were never very good at explaining these purely psychological aspects of language use ("RT and error data"). And Chomsky notoriously dismissed such considerations under the head of "linguistic performance" so that, despite his allegedly psychological view of grammar, linguistic theories were ring-fenced from refutation by genuinely psychological data. Connectionist systems, because they do model psychological processes, can explain complex patterns of RT and error data. Interestingly, although they follow rules, they do not always clearly have internal structures corresponding to those rules.
7. What of the evidence that Markman & Dietrich cite for cognitive systems having rules embedded in them? They mention classical AI reasoning systems (paragraph 56), but say little more than that these can be described as following rules, which is not at issue. Similarly (paragraph 58) they state that psycholinguists have typically assumed "mental representations of... syntactic structures", but they do not say what arguments justify such an assumption. Psycholinguistics also provides the arguments in paragraph 62. But the work of Pinker, Prince and associates merely shows that certain types of rule are needed to describe facts about past tenses. Notably (see paragraph 1, above), Pinker and Prince (1988) claim that the associationist description of past tense inherent in connectionist systems is not adequate to model the intricacies of real past tense systems.
8. Thus the claim by Markman & Dietrich that "there have been some impressive demonstrations of rule use" (paragraph 62) is unsubstantiated. Rather, Pinker's work shows that people follow one set of rules and not another. The sense in which "these findings support the view that grammar is mediated by rule-governed processes" (paragraph 62) is that the past tense system must be described by classical grammatical rules and not the statistical processes of connectionism. All it says about the internal mechanism is that it must in some sense match the complexity of the grammatical rules. Pinker and Prince claim that connectionist mechanisms cannot do so.
9. Markman & Dietrich also mention developmental sequences, such as those found with balance beam problems, as support for the use of internal rules. However, the fact that different rules are followed at different stages no more proves that the cognitive system has internal structure corresponding to those rules than does the fact that children at a single stage can be described as following a single rule.
10. The arguments that Markman & Dietrich put forward to suggest that some behaviors are not inherently rule-governed suffer from the same flaw as those they put forward in favor of inherent rules. They show that one rule system is inadequate to describe a certain behavior, without providing evidence against the (inherent) use of a different rules system. In particular, the locomotion example (Thelen and Smith, 1994) discussed in paragraphs 59 and 60 shows only that behavior sometimes gets described wrongly.
11. Nothing in what Markman & Dietrich say allows the conclusion that mediating states are rule-governed in the sense that "rules... are part of the cognitive system" (paragraph 62). Indeed it is not obvious that there are any convincing arguments for this conclusion, though cases where people explicitly follow rules (e.g., chess playing) may need to be considered separately. So, what follows from the fact that mediating states are not necessarily direct representations of the rules that a cognitive system is following? The gains are mainly conceptual rather than empirical. The idea of "rules in the mind" has always been a problematic one, since the notion of a rule presupposes a sentient being who may or may not follow the rule. Wittgenstein, who devoted much effort to explicating the notions of rules and rule following, claimed (1958, Part II, paragraph xiv) that "in psychology there are experimental methods and conceptual confusion". The question of whether rules are inherent in cognitive mechanisms seems to be one such area of conceptual confusion. Furthermore, arguments that appear to be directed against the whole of generative linguistics and cognitive science, many of which have come from Wittgenstein scholars (e.g. Baker & Hacker, 1984), lose much of their validity if the notion of the mind containing rules, rather than its operation being described by them, is given up.
12. Giving up this idea does not, however, mean giving up the main tenets of cognitive science. We still need to describe accurately the sentences that people use, and to explain which structures they have difficulty with. Thus, for example, even though connectionist mechanisms, unlike classical AI mechanisms, are reasonably successful in accounting for genuinely psychological data, if they cannot support the behavior that people exhibit when they have learned, say, the past tense system of a natural language, we need to look elsewhere for a model of the mechanism that supports language behavior.
13. Neither does giving up on the idea of rules in the mind mean embracing Platonism, a route taken by some generative linguists, most notably Katz (1981). Just as Platonism in the philosophy of mathematics is not inevitable, neither is Platonism in the philosophy of language (see, Garnham, 1990).
Baker, G. P., & Hacker, P. M. S. (1984). Language, sense and nonsense: A critical investigation into modern theories of language. Oxford: Blackwell.
Chomsky, N. (1980) Rules and representations. Behavioral and Brain Sciences 3 : 1-61.
Fodor, J. A. (1965). Could meaning an Rm? Journal of Verbal Learning and Verbal Behavior, 4, 72-81.
Garnham, A. (1990). Don't ask Plato about the Emperor's Mind: Comment on "The emperor's new mind" by Roger Penrose. Behavioral and Brain Sciences, 13, 664-665.
Guthrie, E. R. (1935). The psychology of learning. New York: Harper.
Katz, J. J. (1981). Language and other abstract objects. Totowa: Rowman and Littlefield.
Markman, A.B. & Dietrich, E. (1998). In Defense of Representation as Mediation," PSYCOLOQUY 98 (9). ftp://ftp.princeton.edu/pub/harnad/Psycoloquy/1998.volume.9/psyc.98.9.48.representation-mediation.1.markman
Mowrer, O. H. (1960). Learning theory and the symbolic processes. New York: Wiley.
Osgood, C. E. (1953). Method and theory in experimental psychology. New York: Oxford University Press.
Pinker, S & Prince, A. (1988) On language and connectionism: Analysis of a parallel distributed processing model of language acquisition. Cognition 28(1-2): 73-193.
Thelen, E., & Smith, L. B. (1994). A Dynamic Systems Approach to the Development of Cognition and Action. Cambridge, MA: The MIT Press.
Wittgenstein, L. (1958). Philosophical investigations (2nd Ed.). Oxford: Blackwell.