High-speed simulation of the dynamic neural responses of retinal and cortical simple neurons to complex visual scenes using general purpose computing on graphics processing units.

The spatiotemporal properties of single-cell responses to light have been investigated using simple visual stimuli in physiological experiments, and models of neuronal responses have been constructed on the basis of these experiments. However, incoming visual images from natural environments to the retina are spatiotemporally more complex. Therefore, these models are not enough to fully explain the visual functions of retinal and cortical neurons. To understand the visual functions of neurons in the neural visual processing system, the two-dimensional responses of neuronal models to time-varying scenes from natural environments should be studied through real-time observations.
    The retinal bipolar and amacrine cells exhibit a center-surround antagonistic spatial receptive fields but they exhibit different temporal properties in responses to light, known as sustained and transient responses, respectively. The cortical simple cells have a Gabor-function spatial receptive fields with a temporal receptive fields varying over hundreds of milliseconds. These spatial properties of receptive fields are modeled as a spatial convolution with a large two-dimensional digital filter kernel, which requires large computational costs. Additionally, these temporal properties of receptive fields are often implemented as a temporal convolution using a lookup table,which also requires large computational costs. Therefore, the size and operation speed of a model are often limited by a finite amount of available computational power at a time of a conventional digital computing system based on sequential processing.
     A general purpose computing on graphics processing units is a highly effective approach for speeding up simulations of visual processing systems. In the present study, we have developed a high-speed simulator running on a single GPU which reconstructed the corresponding neural images formed by retinal and cortical simple neurons with physiologically reasonable spatiotemporal properties by virtue of the efficacy of parallel processing of GPU. Here, we conducted simulations for computer-generated, time-varying visual scenes with 256 x 256 sustained and transient retinal neurons and eight sets of 128 x 128 cortical simple neurons of separable and non-separable space-time receptive fields of different preferred orientations at 80 frames per second using a NVIDIA Quadro 4000 for Mac.