With the advent of ultra-high field magnetic resonance imaging (UHF MRI), structural and functional features of the human cerebral cortex can be studied at a sub-millimeter ‘mesosopic’ scale beyond the common resolution of standard MRI scanners. At this spatial resolution it becomes possible to differentiate brain activation in different cortical layers and columnar clusters coding sub-categorical features. I will present recent experiments that show that it is possible to topographically map feature representations in specialised brain areas and to relate mapped columnar clusters in the human motion complex (hMT+) to consciously perceived directions of motion in an ambiguous motion display. I will also describe how the high signal-to-noise of 7T fMRI combined with denoising autoencoder networks recently allowed us to decode imagined letter shapes from retinotopic activity patterns in early visual cortex. Finally I will outline how these and other findings are currently integrated in a biologically inspired embodied computational architecture simulating perception and action in a virtual 3D environment.