A Simple Algorithm for Topographic ICA.

A number of algorithms have been proposed which find structures that resembles that of the visual cortex. However, most of the works require sophisticated computations and lack a rule for how the structure arises. This work presents an unsupervised model for finding topographic organization with a very easy and local learning algorithm. Using a simple rule in the algorithm, we can anticipate which kind of structure will result. When applied to natural images, this model yields an efficient code for natural images and the emergence of simple-cell-like receptive fields. Moreover, we conclude that the local interactions in spatially distributed systems and local optimization with norm L2 are sufficient to create sparse basis, which normally requires higher order statistics.