next up previous
Next: Processing Up: Playpen Previous: Playpen

Architecture

Playpen [Gasser ColungaGasser Colunga1997,Gasser ColungaGasser Colunga1998] is a connectionist model of the acquisition of word meaning. The network itself is a generalization of a continuous Hopfield network; that is, it consists of symmetrically connected simple processing units which respond by updating their activations when certain units are ``clamped'' to a particular input pattern. The long-term knowledge in the network is encoded in the weights on the connections; these are either hard-wired or learned as the network is presented with training patterns.

We divide the units into three layers (Figure 17) a VISION layer, a SPATIAL CONCEPTS layer, and a WORDS layer. In our work to date, we have treated the VISION and WORDS layers as input/output layers, and the SPATIAL CONCEPTS layer as a hidden layer. That is, when we present a pattern to the network for training or testing, it is the VISION or WORDS units which are clamped to the values in that pattern, and then it is the VISION and WORDS units which we treat as the network's response to the input pattern. The SPATIAL CONCEPTS units are hidden in the sense that they are not directly accessible ``from the outside'' (though we can of course observe their activations as the network runs).


  
Figure: Basic Playpen architecture. Each rectangle represents a layer of MOUs and/or MRUs and each arrow a pattern of connectivity between layers.
\begin{figure*}\centerline{\psfig{figure=architecture.eps}} \end{figure*}

As discussed above, units in the network are of two basic types. Micro-object units (MOUs) represent primitive features of objects. These are just the familiar processing units in networks of this general type except that in addition to an activation, each has a phase angle. Phase angles provide a solution to the binding problem; activated MOUs with similar phase angles represent features of a single object. MOUs which are in phase with one another affect each other's activation more strongly, units connected by positive weights tend to attract each other's phase angles, and units connected by negative weights tend to repel each other's phase angles.

Micro-relation units (MRUs) represent features of relations. While each MRU is a single unit in the sense that it has a single activation, it has two separate micro-roles, and separate connections from each micro-role to other MOUs and MRUs. Each micro-role has its own phase angle, corresponding to the phase angle of one of the objects that is being related. All else being equal, an MRU tends to be activated to the extent that it receives inputs on its two micro-roles which are maximally out-of-phase, that is, inputs representing two distinct objects.

Relational correlations take the form of connections between MRUs. Each connection is really a pair of connections, mapping the micro-roles of one MRU to those of the other. Positive connections cause one MRU to tend to activate the other and align its phase angles with it. Negative connections cause one MRU to tend to inhibit the other.


next up previous
Next: Processing Up: Playpen Previous: Playpen
Michael Gasser
1999-09-08