The basic execution cycle of the performance model consists of three steps implementing a constraint propagation process:
The cycle is repeated until the data registers reach a quiescent state.
A constraint element is excited if its excitation strength exceeds a certain threshold. The excitation strength is measured by the Hamming distance between the classifier of the constraint element and the bit patterns in the data registers. Multiple competing constraint elements can be excited at any instant. When an excited constraint element is activated, it gains exclusive control over the data registers, preventing other constraint elements from writing over the register contents. As the register contents change, an activated constraint element might be deactivated and relinquish its control.
The constraint propagation process is not committed to using any particular classifier in a predetermined way. A classifier may use partial semantic information to enforce constraints on the phoneme register. It may also use partial phonological information to infer semantic information. The propagation process can be freely intermixed with the addition of new constraints and modification of the existing ones.
To illustrate how competing constraint elements can cooperate to enforce phonological and semantic constraints, we examine a simple situation. Assume that at some time the meaning identifier describing a red, round, edible fruit appears in the meaning register. These bits might have been set by a vision module that has recognized an apple or a picture of an apple, or perhaps by an olfactory module that has recognized the smell of an apple. We also assume that for some reason the plural bit of the grammar register is set, perhaps because the picture shows two apples.
Suppose also that at this point the performance model has two classifiers: a rote-classifier for the apple constraint, which captures the correlation between the phoneme sequence [ae.p.l], the [+red +round +edible +fruit] meaning, and the [+noun -verb -plural ...] grammar, and a rule-classifier for the voiced-plural rule, which captures the phonological rule that the plural of a noun ending in a voiced phoneme is formed by appending a [z] phoneme to the noun.
The situation at the initial time is depicted in Figure 4. The initial situation triggers a sequence of classifier actions to fill the slots of the phoneme register with the sound sequence corresponding to the plural of ``apple.'' The content of the meaning register is sufficient to activate the constraint element described by the rote-classifier for apple. The apple constraint then attempts to set as many unknown bits as it can. It asserts the bits describing the phoneme sequence into the phoneme register. This encounters no resistance because all of those bits were initially unknown. The apple constraint also sets some grammar bits. The noun bit is turned on and the verb bit is turned off. However, a conflict arises over the setting of the plural bit. The picture of two apples forced the plural bit on, but the apple constraint is trying to assert a singular. Figure 4(b) shows the contents of the registers at this point.
Figure 4: Generating sound from meaning. (a)
Initial state: The performance model has two classifiers: a
rote-classifier for the apple constraint and a rule-classifier for the
voiced-plural constraint. An event fills the meaning register with
features describing an apple, and the grammar register with the
[+plural] feature.
(b) Apple constraint excited: The apple constraint fires and
writes the sound sequence [ae.p.l] into the phoneme register.
Some unknown grammatical bits (such as ?noun and ?verb) are also
filled by the apple classifier. Note that a conflict arises
over the assignment of the plural bit.
(c) Voiced-plural constraint excited: The
voiced-plural constraint sends a
shift left signal to the phoneme register, and fills the unknown
terminal slot with the [z] phoneme. The voiced-plural constraint also
restores the conflict plural bit to 1. The apple constraint is
deactivated.
(d) Quiescent state: The system reaches a consistent state with
the pronunciation of ``apples,'' [ae.p.l.z], in the phoneme register.
No new constraints are excited.
All the phoneme bits from the apple constraint are now in the phoneme register. The fact that there is a noun under consideration (+noun in the grammar register), that there is a conflict over the plural bit, and that the terminal [l] phoneme is [+voice] is a sufficient trigger to activate the constraint represented by the voiced-plural classifier. It sends a shift left signal to the phoneme register, moving the phonemes ae.p.l to less recent positions, and locking the determined phonemes so that they cannot change. The most recent phoneme slot is filled with unknowns, which are certainly allowed to change. The apple constraint now becomes less excited because the values it would like in the phoneme register are all in conflict with the ones that are there. The voiced-plural constraint now fills the unknowns in the current phoneme slot with the phoneme [z]. See Figure 4(c).
As the apple classifier is deactivated, it drops its attempt to set the plural bit to 0. The noun, the verb, and the plural bits retain their last values. The plural bit is still in conflict, but it will put up no resistance if another constraint tries to turn it on. In particular, the excited voiced-plural rule-classifier restores the plural bit to 1. At this point the system reaches a quiescent state (Figure 4(d)) with a consistent representation of the plural noun pronounced [ae.p.l.z] in the phoneme register.
Constraint elements infer meaning from sound as well as sound from meaning. For example, using the same two classifiers as before, the performance model can fill in the grammatical and semantic details as the sound pattern of ``apples'' is shifted into the phoneme register. The same mechanism of constraint elements and shift registers is effective for both production and comprehension of a word.