Page 98 - Encyclopedia of Philosophy of Language

Language and Mind
cognitive modeling do not perform such simulations. How should such machines be described? On one view they do not contain representations of the rules they appear to be following. Rules are traditionally rep- resented symbolically, and connectionism has been described as a subsymbolic approach to cognition. A contrasting view is that connectionist machines rep- resent rules indirectly, and usually in a distributed fashion. Trying to decide which account is correct is complicated by the fact that many connectionist machines do not exactly follow the rules that their designers wanted them to, so it is not surprising that they do not represent those rules. In one famous exam- ple, a network was trained to produce the past tenses of English verbs from their stem forms (Rumelhart and McClelland 1986). However, a detailed analysis of the performance of the machine (Pinker and Prince 1988) showed its knowledge to be lacking in many respects. In particular, it had not encoded the fact there are no phonological conditions on whether the regular (-ed) rule can be applied. Some connectionist systems do follow (usually much simpler) sets of rules exactly, and properties of their (matrix algebra)
descriptions may correspond to information that one would intuitively want to say is represented in the system. However,this representation is not so obvious as a traditional symbolic one. So connectionist machines raise in an acute form the question of when information is represented explicitly and when implicitly.
The contrast between implicit and explicit rep- resentation can be illustrated with a simpleexample from semantic memory. Sparrows are represented as a subclass of birds. Birds are represented as being able to fly, unless there is specific information to the contrary. There is no specific information that spar- rows cannot fly. From the explicitly encoded infor- mation it can, therefore, be inferred that sparrows can fly—that information is implicitly represented. Whether information is encoded explicitly or implicitly determines how easily a particular task can be performed. Implicit information should take longer to compute than explicit information takes to retrieve. It may appear from this example that the contrast
between explicit and implicit representation is a clear one. However, it is not, as the questions raised by representation in connectionist networks show. Indeed, although the contrast between implicit and explicit representation has become increasingly important recently, it remains unclear whether there is one distinction or several.
Although connectionist machines raise important questions about how mechanisms encode rules and follow them, their existence in no way bears upon the very difficult philosophical questions about rule following raised by Wittgenstein (1953), which have sometimes been taken to challenge Chomsky's (e.g., 1972) idea of linguistic rules in the mind. The descrip- tion of a connection machine (or, for that matter, a von Neumann machine) as following a rule is part of a description of its behavior by us. Wittgenstein's questions about how people follow rules turn into questions about what we, as cognitive scientists expect of a machine that we describe as following a certain set of rules. It does not matter whether those rules are encoded explicitly, or only implicitly.
See also:Intentionality.
Bibliography
Chomsky N 1972 Language and Mind. Harcourt Brace Jovanovich, New York
Dretske F 1988 Explaining Behavior:Reasons in a World of Causes. MIT Press, Cambridge, MA
Fodor J A 1975 The Language of Thought. Crowell, New Y ork
Fodor J A 1990 A Theory of Content and Other Essays. MIT Press/Bradford Books,Cambridge,MA
Marr D 1982 Vision: A Computational Investigation into the Human Representation and Processing of Visual Infor- mation. Freeman, San Francisco, CA
Pinker S, Prince A 1988 On language and connectionism: Analysis of a parallel distributed processing model of language acquisition. Cognition 28:73-193
Rumelhart D E, McClelland J L 1986 On learning the past tenses of English verbs. In: McClelland J L, Rumelhart D E, et al. Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Vol.2: Psychological and Biological Models. MIT Press, Cambridge, MA
Wittgenstein L 1953 Philosophical Investigations (trans. Anscombe G E M). Blackwell, Oxford
1. Statement of the Hypothesis
The relationship between language and culture, or language and world view, has been noted at least since Wilhelm von Humboldt (1836). But discussion
remained relatively dormant until the 'Golden Ageof Native American Indian Linguistics' in the first half of the twentieth century.
Although everyone calls it the Sapir-Whorf hypoth-
76
Sapir-Whorf Hypothesis O. Werner