The simulation was based on an auto-associative network. The functional architecture of the network is illustrated in Fig. 6.3. It contained two pools of 24 “visible” units1 that were used both to present input vectors to the network and to record associated output patterns. There was also a layer of 80 semantic units, used to encode two types of putatively modality-speciﬁc semantic features. One type of unit was dedicated to encoding visual properties of objects, and the other type was used to encode functional properties. There were bidirectional connections among units both between and within pools, but no direct connections between the two pools of “visible” units.