This paper explores an inferential system for recognizing visual patterns. The system is inspired by a recent memoryprediction theory and models the high-level architecture of the human neocortex. The paper describes the hierarchical architecture and recognition performance of this Bayesian model. A number of possibilities are analyzed for bringing the model closer to the theory, making it uniform, scalable, less biased and able to learn a larger variety of images and their transformations. The effect of these modifications on recognition accuracy is explored. We identify and discuss a number of both conceptual and practical challenges to the Bayesian approach as well as missing details in the theory that are needed to design a scalable and universal model.