Vision: Computing efficiently

Retinal prostheses aim to restore vision in patients with damaged photoreceptors by bypassing the retina, the sheet of brain tissue that lines the eyeball. A retinal implant's size, weight, and power consumption must be comparable to the retina's, which is half a millimeter thick, weighs half a gram, and consumes a tenth of a watt.

Retinal microcircuit
Silicon retina's response to a face Its four ganglion cell types (color-coded) spike where there is spatial contrast (red for dark; green for light) or when there is temporal change (blue for increase; yellow for decrease). The reconstructed video is shown above. [Kareem Zaghloul 2004]

Existing prosthesis designs simply stimulate when incident light, captured by a video camera, exceeds a threshold.

The retina’s ganglion cells, whose signals the optic nerve carries, interpret a small portion of the visual scene for the brain. They compare the signals of half a dozen to several hundred photoreceptors. They spike when the overall intensity changes, or when the central region differs from the surrounding one. It is the percentage change over time or space—rather than the absolute change—that matters; an adaptation that allows us to see in dim starlight as well as bright sunlight.

We have replicated the ganglion cells' spiking responses in an artificial retina by emulating the retina's neural circuits with transistor circuits.

It is feasible to implant our silicon retina inside the eye as it dissipates only 60 milliwatts—a thousandth of what a computer consumes. When mounted flat against the retina, the chip could be wired to stimulate healthy ganglion cells, which lie near the surface, allowing a blind patient to see again. We are currently duplicating the cortex’s visual perception in silicon.

Kareem Zaghloul's silicon retina includes four ganglion-cell types.
Paul Merolla's visual cortex chip has orientation maps.

Peter Sterling
Jon Demb

Whitaker Foundation
Packard Foundation's Interdisciplinary Science Program