An Architecture for the Learning of Perceptually Grounded Word Meanings

In this statement, we discuss two kinds of properties that a grounded model of the learning of word meaning should have, those related to the way in which linguistic and non-linguistic processing should interact and those related to the representational demands placed on such a model. We also introduce Playpen (Gasser & Colunga 1997) a neural network architecture with these properties. Playpen is a generalization of a continuous Hopfield network in which processing units have phase angles as well as activations and units interact through the attraction and repulsion of their phase angles as well as the usual excitation and inhibition. Phase angles allow the network to bind the features of multiple, simultaneous objects. In addition to units representing features of objects, Playpen also has micro-relation units (MRUs) representing relation features. Each MRU has two phase angles, one for each of the objects it relates. As in other attractor neural networks, processing consists of pattern completion: certain units in the network are clamped and others are allowed to take on values that are consistent with the clamped values. Though the network is divided into layers of units representing WORDS, CONCEPTS, and VISUAL PROCESSING, none of these layers has any particular priority in processing. Learning in the network makes use of a variant of contrastive Hebbian learning (MoveUan 1990). Again no units have special priority; any may be treated as input or target with respect to the learning algorithm.