Neural network simulation of a discrete model of continuous effects of irrelevant stimuli

Stochastic cascade processes as a model of multi-stage concurrent information processing

McClelland's (Psychol. Rev. 86 (1979) 287) influential cascade model of reaction time (RT) is often seen as an important explicitly formulated alternative to strictly serial models because some of its basic notions fit in with known characteristics of neural activation processes. A disadvantage is that it is an essentially deterministic model to which noise is added by across-trials parameter variability that is unrelated to the concept of cascaded activation propagation per se. We propose and analyze a general stochastic cascade model based on neural counting processes; within this framework we present a new process-oriented interpretation of McClelland's cascade activation equation. We then theoretically explain and numerically explore conditions under which the stochastic cascade model predicts additive vs. interactive mean RT effects, and compare them to known properties of McClelland's model. Some of these properties remain valid even within a stochastic neural counting formulation, while others reflect assumptions which are essentially unrelated to the concept of cascaded activation.

Testing the global-slowing hypothesis: are alcohol's effects on human performance process-specific or task-general?

In an interesting recent meta-analysis, Maylor and Rabbitt (1993) suggested that alcohol's effects on human performance may not be process- or stage-specific, but reflect a general, undifferentiated, cognitive slowing. According to this view, performance is globally slowed by a constant multiplicative fraction (b), such that the longer a process takes without alcohol on board (a -), the more it will be slowed by alcohol (a +). In summary: RTa+ = b b (RTa-). In this sense, the effects of alcohol are determined simply by the duration of a process or stage--not by its function or content--and attempts to map the effects of alcohol to specific cognitive operations are essentially futile. This global-slowing hypothesis entails, then, (i) that the function relating RTa+ to RTa- will be linear and increasing; (ii) that the value of b will be significantly greater than 1.0; and (iii) that all experimental factors which increase the complexity (hence, duration) of a task or stage will interact with alcohol. In this study we tested the global-slowing hypothesis directly using fixed set, varied set and concurrent sets item-recognition paradigms. All three tasks showed convincing additivity between alcohol and other key experimental factors which affect response latency (e.g., setsize, response type); there was no hint of any of the spectrum of significant interactions predicted by the global-slowing hypothesis. A meta-analysis of varied set latencies, analogous to Maylor and Rabbitt's, yielded a reasonably linear alcohol/no-alcohol function, but with a slope constant (b) less than 1.0. In all, the data provided little support for the global-slowing hypothesis.

Overlapping stage models and reaction time additivity: effects of the activation equation

Many experimental factors have been found to affect mean reaction time (RT) additively in factorial experiments. What sorts of RT models are compatible with this fact? Sternberg (1969) showed that serial, discrete-stage models are consistent with additivity, and as a result additivity has sometimes been regarded as evidence in favor of such models. However, McClelland (1979) showed that an alternative "cascade" model, which violates crucial assumptions of discrete-stage models, also predicts RT additivity in many cases. This article examines various modified versions of the cascade model, and shows that mean RT additivity arises from many overlapping stage models, including some with thresholds, information quantization, and nonlinear activation transformations. This suggests that other aspects of the data besides mean RT additivity should be examined to distinguish between serial and overlapping stage models.