Speculations on Leveraging Graphical Models for Architectural Integration of Visual Representation and Reasoning