Two doorways into the visual system
Millions of people are blind not because the brain's visual machinery is gone but because the front end has failed — photoreceptors lost to retinitis pigmentosa or age-related macular degeneration, or an eye and optic nerve destroyed by trauma or glaucoma. Visual neuroprostheses try to re-enter that intact machinery through one of two doorways. A retinal prosthesis leaves the eye in place and stimulates the surviving inner retina, borrowing the eye's optics and its orderly wiring. A cortical visual prosthesis bypasses the eye and optic nerve entirely and writes directly onto primary visual cortex. Which doorway you choose determines everything downstream — how much of the visual code you must recreate, how much surgical risk you accept, and which patients you can even help.
You reach this track already fluent in motor decoding: read intention off cortex and drive a cursor or arm. Vision inverts that entire premise. Here you are not decoding — you are encoding, injecting patterned activity that the brain will interpret as a scene. The device is the source, the brain is the receiver, and success is measured not by classification accuracy but by what the person reports seeing. That inversion, more than any hardware limit, is what makes artificial vision the hard problem it is.
The phosphene: the atom of artificial vision
Every electrical visual prosthesis is built from one primitive: the phosphene, a localized spot of light evoked when you stimulate a patch of retina or cortex. A phosphene is to artificial vision what a pixel is to a screen — but the analogy flatters the device. A real phosphene is fuzzy rather than square, varies in size and brightness from electrode to electrode, may appear as a blob or a streak or a cloud, drifts in apparent position, and fades within seconds if you hold the stimulus constant.
The naive engineering model says: N electrodes give an N-pixel image, so scale the array and the picture sharpens. Reality refuses this. Neighbouring phosphenes overlap and merge because injected current spreads; some electrodes evoke nothing; others evoke distorted shapes; and the phosphenes do not tile the visual field on a neat grid. The effective resolution is far below the electrode count, and closing that gap is the substance of this entire track.
Why more electrodes is not the answer
In motor BCI, channel count scales performance almost lawfully. Vision is different: the bottleneck is not how many electrodes you can pack but how faithfully you can encode a scene and how the brain interprets what you write. A first, concrete reason is that the map from visual field to cortex is profoundly non-uniform. Cortical magnification means a degree of visual angle near the fovea occupies far more cortical millimetres than a degree in the periphery.
The cortical magnification factor M (cortical mm per degree of visual angle) falls steeply with eccentricity E; M_0 and E_2 are fit constants. A uniform electrode grid on cortex therefore samples the visual field very non-uniformly — dense at the centre of gaze, sparse toward the edges.
The consequence is unavoidable: a regular array lands on a warped slice of visual space, so the phosphene map the patient perceives is itself warped. Any honest encoder must model this geometry and pre-distort the stimulus to compensate — you cannot treat the electrode grid as a Cartesian screen. This is one reason raw channel count buys less here than intuition suggests.
Where the field actually stands
An honest read separates what is demonstrated from what is aspired to. Demonstrated: blind volunteers perceive phosphenes, localize them in space, detect motion and direction, and read isolated large letters slowly; and in 2021 a single patient with retinitis pigmentosa recovered partial, goggle-mediated vision by an optogenetic route — the first such report in a human. Emerging: photovoltaic subretinal implants have let patients with atrophic macular degeneration read letters and words in trials. Open: no device yet restores fluent reading, face recognition, or anything close to a naturalistic visual world.