In four experiments, participants’ performance on a variety of nonlinear patterning discriminations was assessed using a predictive learning task and visual patterns. Between groups, the similarity of the stimuli that composed these visual patterns was manipulated. When the stimuli were of low similarity, participants’ performance was consistent with the predictions of one version of Pearce’s (1987, 1994, 2002) configural theory of learning (Kinder & Lachnit, 2003); they were better able to discriminate between different patterns when they shared few, rather than many, stimuli. This effect was not observed when the similarity of the stimuli was high. Under these conditions, the results were more consistent with predictions of elemental theories of learning (e.g., Rescorla & Wagner, 1972; Wagner, 2003). This is the first time that these different patterns of performance on complex patterning discriminations have been shown within a single stimulus modality in the same experiment. The overall pattern of results is difficult to reconcile with either elemental or configural models of associative learning.