Researchers at EPFL say their AI models can predict exactly which patterns of brain stimulation will make a user "see" an object — a step beyond today's vision prosthetics, which can only evoke abstract spots of light called phosphenes.
From phosphenes to faces
Today's cortical visual prostheses work by stimulating electrodes on the visual cortex; recipients perceive small bright dots that the brain has to assemble into anything resembling a scene. The EPFL NeuroAI Lab, led by Martin Schrimpf, trained brain-aligned vision models to predict which combinations of electrodes evoke a recognizable face or object — not just brighter or dimmer phosphenes.
Tested on sighted monkeys at the Donders Institute
Dutch collaborators at the Donders Institute used the EPFL models to drive live cortical stimulation in sighted macaques. Preliminary results, presented at the International Conference on Learning Representations (ICLR 2026) in April, showed the model-chosen stimulation patterns reliably influenced the animals' behavior on visual-object recognition tasks. "Our model turned out to be quite efficient in predicting which stimulation pattern would yield a strong effect on the monkeys' behavior," Schrimpf said in the announcement.
Why this matters for blind patients
Object-level perception, rather than dotted outlines, is the missing ingredient that has held cortical implants back as a serious assistive technology. Pairing model-guided stimulation with the next generation of higher-channel-count implants — the same generation pushing brain and nerve interfaces into commercial prosthetic devices — could finally make functional sight realistic for patients with cortical-route blindness.
A bigger AI-in-medicine moment
The work lands as regulators and surgeons grow more comfortable with AI inside the operating room and the clinic, from FDA-cleared AI imaging to surgical robotics platforms that lean on perception models for tissue tracking.
Reporting based on coverage from EPFL News, TechXplore and Mirage News covering NeuroAI Lab work presented at ICLR 2026.