A Variational Neural-Field Interpretation of the Stereo-CeNN Architecture in Light of Disparity-Tuned Cortical Organization

Stereoscopic vision relies on the ability of the visual system to estimate binocular disparity across the visual field. In computational models, stereo matching can be represented through multiple candidate disparity hypotheses defined over retinotopic position. In this work, we reinterpret the Stereo Cellular Neural Network, here referred to as Stereo-CeNN, within a continuous variational framework and introduce a neural-field representation defined over the joint domain of retinotopic position and disparity. In this interpretation, the discrete activities of the Stereo-CeNN are regarded as samples of a continuous activity field r(x,d,t), where different disparity layers correspond to populations tuned to different disparity preferences. The binocular input drives these populations according to local matching evidence, while spatial interactions within each disparity channel promote coherent activity across neighboring retinotopic locations. The final disparity estimate is obtained, consistently with the original Stereo-CeNN architecture, by selecting the disparity channel with the largest steady-state response at each retinotopic position. We show that the layer-wise organization and local spatial dynamics of the Stereo-CeNN admit a continuous neural-field counterpart, establishing an explicit structural and variational correspondence between the discrete retinotopy–disparity lattice and a population-level representation of stereopsis. This formulation provides a conceptual bridge between the independently developed Stereo-CeNN architecture and the mesoscale organization of disparity-selective populations observed in visual cortex, without implying a one-to-one anatomical or physiological correspondence. Because Cellular Neural Networks are locally connected dynamical systems originally conceived for analog and mixed-signal VLSI implementation, the same organization is also compatible with general neuromorphic principles of local, parallel, and distributed computation.