Retinotopic Sparse Representation of Natural Images

Independent component analysis and sparse coding have provided a functional explanations of simple cells in primary visual cortex (V1). The learned components(corresponding to the responses of neurons) of these models are randomly scattered and have no particular order. In practice, however, the arrangement of neurons in VI are ordered in a very specific manner. In this paper, we propose a sparse coding of natural images under a retinotopic map constraint. We investigate the spatial specifically connections between retinal input and v1 neurons. Some simulations on natural images demonstrate that the proposed model can learn a retinotopic sparse representation efficiently.