Abstract
The visual areas V1, V2, and V3 are defined based on a systematic mapping of visual space on to the cortex, following alternating vertical and horizontal meridian representations around a common fovea-to-periphery axis. Why do we have multiple visual areas and why are they organized this way? To gain insight into these questions, we used a self-organizing map algorithm which has successfully predicted both micro-scale pinwheel organizations within V1 and large-scale motor cortex organization (Durbin & Mitchison, 1990; Graziano & Aflalo, 2007). This algorithm takes a high-dimensional space as input and projects it onto a two-dimensional map such that each map-unit has a tuning curve over the input space and adjacent map-units have similar tuning (Kohonen, 1982). To construct an input space, we first discretized the visual hemifield into a high-dimensional space of visual field locations (e.g. 15×30=450 dimensions). Gaussian filters of different sizes were constructed at each location, where each filter corresponds to a point in this high-dimensional space, and the full space is specified in a matrix (filters × locations). Using the self-organizing map algorithm, we found that mapping these multi-scale filter bank input spaces naturally yield the major motifs of the visual field organization: a large-scale eccentricity organization with an upper and lower visual field divide and alternations along the vertical and horizontal meridians. Surprisingly, multiple mirrored visual “areas” emerged in the simulated cortical map without the specification of a hierarchical relationship between features. While early visual areas are often studied separately and their hierarchical relationship is emphasized, this modeling work suggests that their functional role may also be understood together. Broadly, these results provide a computational-level argument that the large-scale organization of early visual field maps can be explained by the simple goal of smoothly mapping a multi-scale filter bank on a two-dimensional cortex.