The objective of experiment 4 is to test whether higher-order correlations can be extracted from the past-tense rules and the plural rules, which have very similar structures.
The generalization algorithm when presented with the past-tense rules (from experiment 3) and the plural rules (from experiment 1) produces a higher-order correlation that says the voicing bits of the ending phoneme of the stem and the affix have to match:
[dc.dc.[-voice].[-voice]] [dc.dc.[+voice].[+voice]]
This higher-order correlation can be interpreted as the voicing assimilation rule described in linguistics books (such as [1]). Voicing assimilation captures cross-categorical generalizations governing the formation of not only plural nouns and past-tense verbs, but also third-person singular verbs, possessive nouns, and several other morphological categories.
Linguists explain complicated phonological processes in terms of the interactions of nearly independent and widely applicable rules. Our learning theory gives a plausible mechanism to produce this kind of compact, elegant phonological rules. The theory also exhibits the ``Waltz'' effect [19] that learning becomes more effective when the learner is exposed to more varieties of constraints.