Some criticisms of connectionist models of human performance

Abstract

Connectionism offers a challenge to current information-processing descriptions of linguistic performance. Upon examination, however, models with the connectionist framework are found to be wrong in important respects or are too powerful to be meaningful. The following observations support these claims. The assumption of interactive activation (i.e., two-way connections between units) of specific connectionist models is shown to be both unnecessary and inconsistent with empirical results. Connectionist models with hidden units are demonstrated to be too powerful; they can simulate different types of results that are generated by different process models. Given the power of connectionist models with hidden units, they can describe results with unrealistic assumptions about the psychophysical relationships that are functional in the task. Connectionist models with hidden units are limited in theoretical value without postulating something like sequential stages of processing in which some categorization occurs before response selection. Notwithstanding these limitations, it is noted that other important properties of connectionism are to be found in existing process models of pattern recognition.