The authors introduce the concept of ‘modular predictive coding’ within the hierarchical coding framework for the visual system by means of an articulated mathematical analysis based on the application of the modular coding scheme. It is verified that natural images are successfully compressed and can be reconstructed by the familiarity neurons. The research aims at verifying how modular predictive coding may be considered a biologically realistic metaphor for the visual system. The authors, after a detailed exploration of data and the presentation of several mathematical elaborations, suggest in their final discussion that the representation of natural visual stimuli by means of predictive coding is not essential and top-down connections are engaged in attention, memory recall and in top-down-gated learning processes, without substantially affecting the fast processing of sensory stimuli.