Cortical encoding of retinal output from natural scenes with sparse representation

The visual cortex of a freshwater turtle, when stimulated by a pattern of light, produces waves of activity that have been recorded experimentally and simulated using a model cortex. It is believed that these activity waves encode features of the visual scene, viz. position and velocity of targets. The goal of this paper is to explore how to estimate target velocity using the activity pattern in the model cortex. We consider five natural video scenes and represent them using sparse, over-complete set of basis functions. The associated coefficients are KL-decomposed to provide appropriate cortical signals. The signals are fed as input to a model of the visual cortex and the associated cortical response of a large number of pyramidal cells are generated. Finally, the cortical response has been displayed as a spatiotemporal signal. The paper concludes with a sketch of an outline as to how the motion field of the input visual scene could be reconstructed from the activities of the cortical cells with two steps of processes: estimation of the conductance patterns of pyramidal cells from the activities of the pyramidal cells and estimation of motion field from the estimated conductance patterns of pyramidal cells.