Proceedings of the 14th Annual Conference of the Cognitive Science Society, 1992,

pp.653-658. Hillsdale, N.J.: Erlbaum.

Taking connectionism seriously:

the vague promise of subsymbolism and an alternative.

Paul F.M.J.Verschure
AI lab, Institute for Informatics, University of Z?rich, Winterthurerstrasse 190, CH- 8057 Z?rich, Switzerland.
e-mail: [email protected]

Abstract

Connectionism is drawing much attention as a new paradigm for cognitive science. An important objective of connectionism has become the definition of a subsymbolic bridge between the mind and the brain.
By analyzing an important example of this subsymbolic approach, NETtalk, I will show that this type of connectionism does not fulfil its promises and is applying new techniques in a symbolic approach.
It is shown that connectionist models can only become part of such a new approach when they are embedded in an alternative conceptual framework where the emphasis is not placed upon what knowledge a system must posses to be able to accomplish a task but on how a system can develop this knowledge through its interaction with the environment.

Introduction

Connectionism has been gaining much attention in cognitive science. On of the reasons is that problems of the traditional cognitivistic approach, like the need for noise and fault tolerance and the capability to generalize, are solvable with connectionist, brain-like, techniques.
This proposal makes the problem of complete reduction (PCR) (Haugeland, 1978), or of how a symbolic description of cognition can be reduced to a non-symbolic one, again highly relevant. In the traditional cognitivistic view cognition is seen as formal symbol manipulation. The basic steps of this approach can be defined as: ?1, Characterize the situation in terms of identifiable objects with well defined properties. 2, Find general rules that apply to situations in terms of those objects and properties. 3, Apply the rules to the situation of concern, drawing conclusions about what should be done.? (Winograd and Flores, 1986, p.15).
The physical symbol system hypothesis (Newell, 1980) can be taken as the most influential formulation of this approach. The hypothesis states

that a physical symbol system (PSS) constitutes the necessary and sufficient conditions for general intelligence. A PSS consists of a set of actions and is embedded in a world that consists of discrete states; objects and their relations. Moreover, a PSS has a "body of knowledge" that specifies the relations between the events in the world and the actions of the system, we can also refer to this body of knowledge as a world model built up with symbolic representations. The actions of the system, either in the world, or internal inferences, are organized around the goals of the system according to the principle of rationality: roughly a system will use its knowledge to reach its goals. An important implication of this conceptualization of cognition is that it can (and must) be modelled at the abstract level of symbol manipulation. The specifics of the implementation are, therefore, of no importance. PCR is no longer an issue since the non symbolic level of brain dynamics is not taken to be very relevant in explaining cognition. The hypothesis of physical symbol systems is often seen as the only plausible model for general intelligence which has no serious competitors (e.g. Pylyshyn, 1989). Despite this claim this paradigm also confronts some serious problems. One of these problems is the symbol grounding problem (Harnad, 1990), or the question of how symbols acquire their meaning. In the cognitivistic tradition the meaning of symbols is taken as given (Newell, 1981), which implies that cognitivism has to resort to a nativistic position: that the "body of knowledge" is just present from the start on. Moreover, one has to assume that the system possesses very reliable transduction functions that allow the coupling between events and objects in the world and their internal symbolic representation. These assumption have been criticized on several grounds. For instance, the genome does not have the coding capacity to represent this body of knowledge (Edelman, 1987), or it still needs to be explained how during evolution this "body of knowledge" could have been acquired (Piaget in Piatelli-Palmarini, 1980). Moreover, practical applications developed within this paradigm, for instance robot control architectures, have not been