One question about which there is little consensus among connectionist models of reaction time, is the question of the locus of inhibition between alternatives. Are inhibitory mechanisms active within each cognitive process, or between them?
According to the within-process inhibition view, the activation of one mental code in a particular process (i.e. within a module) will inhibit all of the other mental codes in that process (i.e. within the same module). According to the between-process inhibition view, activation of a mental code in one process (e.g. stimulus identification) will directly inhibit non-corresponding codes in the following process (e.g. response selection).
Most of the earliest connectionist network models, including McClelland’s (1979) Cascade model, were feed-forward networks (see Rumelhart, McCelland, et al., 1986). This meant that any given unit could only supply input to units in later processes, and there were no connections at all between units within the same module. As a result, these network models naturally had to implement between-process inhibition: each stimulus unit had a positive association with one or more response units, and negative associations with all of the rest of the response units.
This feed-forward architecture also arises from many “learning algorithms”: processes through which the association strengths in a network can change from trial to trial, “learning” based on experience. For example, the model of the Stroop effect by Cohen, Dunbar and McClelland (1990) assigns association strengths based on a standard learning algorithm (known as backpropagation) that “trains” the network on the process of reading (giving color word responses to word inputs) and on the process of color-naming (giving color word responses to color inputs). The result of this algorithm is that, for example, the stimulus unit for the word “red” has a positive association with the response unit for the word “red,” and also a negative association with the response unit for the word “green”. Similarly, the stimulus unit for the word “green” has a positive association with the response unit for the word “green”, and also a negative association with the response unit for the word “red”. The result is between-process inhibition: negative associations connect stimulus units to response units.
McClelland and Rumelhart (1981), however, introduced a framework with a different assumption about connections between units. They describe a semantic activation model with a feature module (with units representing mental codes for individual visual features), a letter module (with units representing mental codes for individual letters), and a word module (with units representing mental codes for whole words). They suggest that units are connected with positive or negative associations based on their consistency: that is, because the existence of the word “THE” is consistent with the existence of the letter “T” in the initial position, the initial “T” letter unit and the “THE” word unit have a positive association; on the other hand, because the existence of the word “ARE” is inconsistent with the existence of the letter “T” in the initial position, the initial “T” letter unit and the “ARE” word unit have a negative association.
One result of this kind of assumption is that all of the units within the same module are mutually inhibitory: the existence of a letter “T” in the initial position of a word is inconsistent with there being any other letter in that initial position, the existence of the word “ARE” is inconsistent with there being any other word, and so on. As a result, this model implements within-process inhibition: activation of any unit within a module (process) inhibits activation of all of the other units in the same module. However, this model also implements between-process inhibition, because the same rules apply to connections between units in different modules (e.g. letter units and word units).
This set of assumptions is used as the basis for the Stroop model developed by Phaf, van der Heijden, and Hudson (1990). In this model, color units (representing alternative possible color codes) inhibited both each other (within-process inhibition) and inconsistent color word responses units (between-process inhibition). Phaf et al. (1990) further argue in support of within-process inhibition by citing neurophysiological evidence: they claim that the neurophysiological phenomenon of “lateral inhibition” (inhibition among all of the nearby neurons in the same layer in cortical tissue) can be thought of as evidence that the brain implements within-process inhibition.
It was soon noted, of course, that having both of these kinds of inhibition is computationally redundant. Several years later, McClelland (1993) proposed a normative framework for connectionist network models of performance called the Graded Random And Interactive Network (GRAIN) framework, in which he suggested that models should use only inhibitory connections between units in the same module, and only excitatory connections between units in different modules. This effectively called for models of performance to constrain themselves to implementing within-module inhibition. Most of the motivation for this move was computational, not psychological, and almost every “advantage” described by McClelland (1993, pp. 659-660) for within-process inhibition could also be found in an appropriately structured model with between-process inhibition.
Many models of consistency effects, nonetheless, have followed this normative suggestion (e.g. Barber & O’Leary, 1997; Cohen & Huston, 1994; Cohen, Servan-Schreiber, & McClelland, 1992; O’Leary & Barber, 1993; Zhang & Kornblum, 1998; Zhang, Zhang, & Kornblum, 1999; Zorzi & Umilta, 1995). Zorzi and Umilta (1995), however, did explore the performance of an alternative version of their model that implemented between-process inhibition. They found that both models could account for performance equally well, and that in the end the only motivation for preferring within-process inhibition was theoretical consistency: because everyone else was doing it.
Kornblum et al. (1999) pointed out that having within-process inhibition can lead to explosive inhibitory feedback effects if activation values are allowed to go below zero (see Kornblum et al., 1999, p. 706). Zorzi and Umilta (1995) and Cohen and Huston (1994) dealt with this by constraining the output of units between 0 and 1. Instead of adding this additional constraint on the dynamics of processing, Kornblum et al. (1999) implement between-process inhibition. By using this alternative inhibition mechanism, and forcing fewer constraints on processing (because output could be either positive or negative with no ill consequences to the model), they were still easily able to account for performance in Simon tasks, Stroop-like tasks, and their factorial combinations.