next up previous
Next: Mathematical Details of the Up: Playpen: Toward an Architecture Previous: Learning

Conclusions and Future Work

In the first part of this report we tried to convince you that in order to understand where the meanings of words come from, it makes sense to look outside as well as within language, especially at human vision and at pre-linguistic conceptual development. But how is one to take on the complexity of these diverse domains, each of which occupies an entire research community? We believe that these research communities have given us much to go on, that we can take enough off the shelf to get us started, and that by slicing our domain of interest very thin, specifically by confining ourselves to simple static spatial relations, it is possible to build an insightful model of the development of word meaning.

In the second part of the report we outlined the beginning of the Playpen project, whose goal is such a model. We described a simple neural network architecture based on features of the human vision system which allows for the emergence of spatial concepts from the interaction of vision and language. We emphasized three basic building blocks which we believe are required: object units, relation units, and delay connections.

In a subsequent technical report (available in September 1997), we describe how Playpen models the relative difficulty of the acquisition of the words on, under, left, and right.

Future work includes

  1. implementation of other components of the architecture, especially those in the What system
  2. application of the model to other specific relations, especially IN and BEHIND.
  3. taking the visual world seriously, eventually using input from a camera.

eliana colunga-leal
Mon Jun 23 04:27:19 EST 1997