Reconstructing Physical Symbol Systems

In attempting to force ALVINN’ into their already bulging symbolist tent, Vera and Simon (1993a, 1993b, 1993c) have burst the seams of the physical symbol system hypothesis (PSSH; Newell, 1980a; Newell & Simon, 1976). We would like to present another view of representation in connectionist networks and respond to statements about ALVINN by both Vera and Simon and Green0 and Moore (1993). In so doing, we will reconstruct the PSSH. The primary error in Vera and Simon’s (1993a) argument is to mistake signal for symbol: “We call patterns symbols when they can designate or denote” (p. 9). This leaves no possibility of nonsymbolic representations. On this view, patterns of neuronal activity in the retina (1993c, p. 125), and presumably even the elementary signaling conventions of doorbells and vending machines (Clancey, 1993, p. 89) are elevated to symbol systems, rendering the concept of “symbol” meaningless. Vera and Simon (1993b) also referred to ALVINN in particular as a “connectionist symbol-processing system” (p. 79), and noted that it is simulated by a von Neumann symbol processor. This conflates what should be two distinct levels of representation, and leads to the mistaken conclusion that any implementation of a Turing-computable function is a symbol processor. We believe that symbol processors are a very special type of device; this is what gives the PSSH significance as a theory of intelligence.