In this chapter we have adopted a very different perspective on the relationship between language and spatial cognition than have the other authors in this book. We have focused on the emergence of relations and language rather than the end-state. In doing so, we have been concerned with the very nature of relations and have presented an account of the ``stuff'' out of which relations are constructed. We believe that this groundwork is essential if we are to understand the full-fledged adult system because we believe the adult system is the product of more of the same sort of process we have proposed for the early emergence of relations, the learning of more and more abstract relational correlations.
In taking this low-level, developmental approach, we have ignored most of the data that are of interest to the other authors in the book, data that have led people such as Jackendoff to posit separate representational modules for spatial concepts, conceptual representations, and language, as well as interface modules connecting them. Thus it should be not be surprising that the picture we offer is considerably simpler than the one offered by others. In a sense our picture is one in which WORDS and VISION are mediated by a single ``interface module'', which we have called SPATIAL CONCEPTS, a layer of micro-object and micro-relation units which is simultaneously under the influence of units in the WORDS and VISION layers.
However, we are aware that the mapping from visual/spatial representations to words and syntax is not a simple one. The SPATIAL CONCEPTS layer in our model is not meant to be taken as a layer of completely connected relation units and object units. Rather we expect this layer in the adult system to consist of multiple sublayers, each performing some form of transformation on the patterns it receives from above or below, each connected to its neighbors by an ``interface'' in the form of a simple pattern of connectivity. As yet we have nothing specific to say about the ultimate internal structure of the SPATIAL CONCEPTS layer nor about how much of this structure is built in and how much develops as spatial categories and words are learned. The point is that our proposals are not incompatible with accounts in which different levels of representation reside in different architectural layers. Our contribution to the question of the putative modules and interfaces joining visual and linguistic representations is twofold. (1) The modularity may be a matter of degree; that is, the representations may simply be more and more linguistic (under the influence of words and syntax) as they appear closer to the strictly linguistic portion of the neural network and more and more visual (under the influence of specifically visual patterns) as they appear closer to the visual portion of the network. (2) Modularity (to whatever degree) can develop, and the interface between language and vision is a place where we might expect this. Therefore, a major goal should be elucidating the mechanisms by which development takes place in response to the visual and linguistic patterns it is exposed to.
Spatial cognition and spatial language are fundamentally relational. In fact, given the importance of space in the young child's world, space may be the place to look for the emergence of relational knowledge. A range of well-attested data are consistent with the view that we have proposed: relational categories, like object categories, take shape as the child picks up on the rich set of correlations available to her in the world and in language, and, like object categories in neural network models, they are built up out of primitive connectionist units.