Abstract

Central to the dispute between symbolic AI and connectionism is the issue of whether cognitive capacities can be or should be explained by the execution of a program. Recently, connectionist models have been developed that seem to exhibit such capacities without executing a program. In my dissertation, I examine one such model and argue that it does execute a program. The argument proceeds by showing that what is essential to running a program is preserving the functional structure of the program. It has generally been assumed that this can only be done by the temporally organized causal capacities of the instantiating system. I observe that counterfactual-preserving functional architecture can be instantiated in other ways, for example geometrically, which are realizable by connectionist networks. I then discuss the consequences of this argument for the role programs play in psychological explanation. The explanation of cognitive capacities via the specification of the functional structure of a program is an instance of functional analysis, and lies at the heart of psychological explanation in symbolic AI. The program specifies steps whose exercise constitutes the exercise of a larger disposition of the system being analyzed. Since the functional structure specified by the program constitutes the analysis of the cognitive capacity, it is the functional structure that does the explanatory work. Because this functional structure can be realized in systems with radically different causal structures, the particular causal structure of a cognitive system is incidental to the analysis of cognitive capacities as such, and any causal structure which realizes the functional structure of a program counts as executing that program. Finally, I explore how this perspective on the relationship between connectionist networks and programs might bear on the issue of learning in connectionist networks. I describe how a dual network architecture might execute a learning program, thus showing how connectionist networks could embody the learning principles of symbolic AI. In this architecture, one network encodes stored knowledge while the other network executes the learning program itself. I then discuss how the interaction of these networks gives rise to the changes in behavior that constitute learning