Unsupervised models of how the brain identifies and categorizes the causes of its sensory input can be divided into two classes: those that minimize the mutual information (i.e., redundancy) among evoked responses and those that minimize the prediction error. Although these models have the same goal, the way that goal is attained, and the functional architectures required, are fundamentally different. This review describes the differences, in the functional anatomy of sensory cortical hierarchies, implied by the two models. We then consider how neuroimaging can be used to disambiguate between them. The key distinction reduces to whether backward connections are employed by the brain to generate a prediction of sensory inputs. To ascertain whether backward influences are evident empirically requires a characterization of functional integration among brain systems. This review summarizes the approaches to measuring functional integration in terms of effective connectivity and proceeds to address the question posed by the theoretical considerations. In short, it will be shown that the conjoint manipulation of bottom-up and top-down inputs to an area can be used to test for interactions between them, in elaborating cortical responses. The conclusion, from these sorts of neuroimaging studies, points to the prevalence of top-down influences and the plausibility of generative models of sensory brain function.