The explanation of the the difference between move and mate has to do with top-down feedback from the word layer to the letter layer. This feedback causes have node sends an excitatory signal to the H node in position one, the A node in position two, the V node in position three and the E node in position four. The net effect of this top-down feedback will depend on which clusters of nodes are active at the word level. There are many words that end in AVE or are of the form MA?E (e.g., made, make, male, mane, mare, mate, maze). By contrast, there are relatively few words that end with OVE or are of the form MO?E (apart from move, the only other word of this form in the model’s vocabulary is mole). Consequently, the top-down feedback to the A node in position two will be much greater than the feedback to the O node in position two, and hence the A node will be more strongly activated. That is, move and mate will not receive equivalent feedforward input from the letter layer. Furthermore, the resonance between the letter and word layers means that the greater activity of the A node will in turn further boost the advantage of words like have and make over the word move, and this in turn will further amplify the differences at the letter layer. McClelland and Rumelhart refer to this as the gang effect. It is another good example of non-linear feedback processes within the model.

Back to Index