Artificial vision encoding model

The transform from camera image to stimulation pattern: the software that decides which electrodes or pixels fire, how strongly, and when, so the evoked phosphenes best convey the scene. Naive encoders simply downsample the image onto the electrode grid; better encoders model the neural code and the perceptual distortions.

Two research directions dominate. Biomimetic encoders reproduce the retina's ganglion-cell code (the retinal encoder of Nirenberg and colleagues), improving reconstructability of natural scenes. End-to-end deep-learning encoders embed a differentiable phosphene model between a neural-network encoder and a task or perceptual loss, optimizing stimulation directly for what the user should perceive or do, and can learn to route around axon streaks and current spread.