Grounding Word Learning in Space and Time

摘要

Abstract This chapter applies dynamic field theory to word learning. The use of one-dimensional neural fields to represent labels and the combination of these with a feature dimension are introduced. These label-feature fields keep a record of prior feature-label associations via the memory trace mechanism. Using a robotic instantiation, the chapter show how individual features of objects, represented in multiple feature-label fields, can be bound via a shared label dimension. The result is a dynamic field model that can 1) learn robust novel label-object mappings after only a few presentations of the label and/or the object, 2) demonstrate emergent categories, 3) fill in missing information, and 4) distinguish between two different objects that share a value on one feature dimension but not others. An expanded version of this model includes two feature-label and two feature-space fields, which enable the model to overcome referential ambiguity by binding names to objects across a shared spatial dimension. This model can capture multiple word-learning behaviors, thus pointing to a critical innovation of this work—the integration of timescales.