Let's return to the plot in which we ask for the five most active nodes (or equivalently in this case, set the activity threshold to 0.05). Not surprisingly, the five word nodes each represent words that are orthographic neighbours of the non-word mave (i.e., they are one letter different). That makes sense: they each have three letters in common with the stimulus. But even though they share the same amount of orthographic overlap with the stimulus, there are large differences in activity between these nodes. Why is that?

A key reason for the activity differences can be seen by zooming in on the activities on the first cycle.

PLOT HERE

As you can see, there are differences in the starting activities of these nodes – for example, have has a head start over save. Differences in the resting activities is the mechanism used in the interactive activation model to capture word frequency effects. Specifically, the resting activities for word nodes are proportional to log word frequency. However, it is also important to note that the size of these activity differences is much greater at cycle 15 (for example) than on the first cycle. This magnification of the initial difference in resting activities is a consequence of lateral inhibition at the word level. Word nodes that start with a higher activity are better able to inhibit competing nodes that have lower resting activities. McClelland and Rumelhart refer to this as the rich-get-richer effect. It is a good example of nonlinear processes at work in the model.

Next: Other factors affecting word layer activity