The preceding sections have indicated that our research on ANNs has produced some promising results for application: faster learning, easier interpretation, better generalization, and medical diagnosis. However, it is crucial to note that our motivations were not to produce such results -- we had no idea at the outset, for example, that the value unit architecture would possess many of the nice algorithmic properties that we have discovered in it. Instead, our motivations are typically started by asking questions about how the biological relevance of ANNs could be enhanced. Indeed, many of our design decisions have been counterintuitive from an "engineering perspective", but have led to unanticipated "engineering" advantages.

Our concern with the relationship between connectionism and cognitive science has been documented in several constructive criticisms (Dawson & Berkeley, 1993; Dawson & Schopflocher, 1992b; Dawson & Shamanski, 1994; Dawson, Shamanski & Medler, 1993). The general theme underlying these criticisms is that much ANN modeling is driven by technological concerns -- building a better mousetrap (i.e., by producing faster algorithms, or architectures that can easily be transferred to silicon chips). Our concern is that very little ANN modeling appears to be motivated by theoretical concerns -- making models that are strongly related to psychological or physiological phenomena. Most of our simulation research has been motivated by these latter concerns, and we hope that we have taken some small steps in enhancing the role of connectionism within cognitive science.

References