Formation and variability of orientation preference maps in visual cortex: an approach based on normalized Gaussian arrays

This work explores formation and variability of orientation preference maps in visual cortex based on normalized Gaussian arrays. An orientation preference map, which has been measured to sketch the orientation preference of neighboring neurons in visual cortex, is emulated by a network of weighted normalized Gaussian arrays. Here the orientation preference map is represented by a set of paired data, each comprising a relative neuron location and the relevant orientation preference, and the mapping structure from the location on a two-dimensional lattice to the orientation preference is then realized by learning a network of normalized Gaussian arrays subject to the paired data. As a result, varieties of mapping structures of different species are essentially characterized by the array number in a network. For example, the icecube structure can be well emulated by using one Gaussian array, the pinwheels structure, that looks like a superposition of two cross icecubes, can be characterized by a network of two Gaussian arrays, and the salt&pepper structure is simply a result of a null Gaussian array, of which the width of each Gaussian unit is set sufficiently large.