As noted above, the model is run on a pattern by clamping (fixing the activations and phase angles of) some of the units on the VISION or WORDS layers. The units in the network are then updated repeatedly -- that is, their activations and phase angles are updated -- until the changes have stabilized. The pattern of activation of the units on the VISION and WORDS layers at this point represents the network's response to the input pattern. Note that the model can be run in both the ``comprehension'' and the ``production'' directions, or even in combinations of the two, by clamping units on different layers.