Grounding computational engines

摘要

Integrating the processing of sensory information and natural language is not a homogeneous enterprise, and there are different proposals from both connectionism and symbolic AI on how to proceed. This paper considers one problematic part of the enterprise, what we call the internalist trap for the systems of symbolic AI and connectionism. Two kinds of computational mechanism are discussed, the Syntactic engine of symbolic AI and the more novel Spatial engine of connectionism, and the different solutions to the internalist trap that each machine requires is explored. What emerges from our discussion is the relative paucity of the representational resources available to the Syntactic engine, in comparison to those available to the Spatial engine. This inequality is important, because it is precisely these resources that, we argue, are crucial in hooking atomic representations to the world. That is, whilst it is possible for both kinds of computational engine to be hooked up to the world, it is only the Spatial engine which possesses the requisite resources in and of itself.