Researchers have taken a significant step toward making “bionic eyes” more predictable by using artificial intelligence to control electrical stimulation in the visual cortex. In a proof-of-concept study involving a blind participant, scientists from UC Santa Barbara, ETH Zurich and Miguel Hernández University used a deep-learning model to design stimulation patterns that produced more accurate and efficient neural responses than conventional approaches. The work could help transform visual cortical prostheses from experimental devices into systems capable of adapting to the unique biology of each user.
The study, published in Neuron, focused on a technology that bypasses the eye and optic nerve entirely. Instead of restoring vision by stimulating the retina, a cortical prosthesis sends electrical signals directly to the visual cortex, the brain region responsible for processing visual information. When electrical current activates groups of neurons in this area, users may experience phosphenes—spots, flashes or shapes of light that can resemble stars, sparks or fireworks. Although these sensations are far from normal vision, they could eventually provide useful visual information to people whose eyes or optic nerves can no longer transmit signals.
The research team tested its approach in a 27-year-old man who had lost his sight after a traumatic brain injury. The participant had a 96-channel electrode array temporarily implanted in his visual cortex at Hospital IMED Elche in Spain. The device was designed both to stimulate neural tissue and to record the brain’s electrical activity. Because the implant was scheduled to be removed after six months, the researchers had a rare opportunity to observe, in real time, how specific stimulation patterns altered activity in the human visual cortex and how those changes corresponded to the participant’s perceptions.
Traditional visual prostheses generally rely on predetermined stimulation rules. Engineers select which electrodes to activate, adjust the amount of current and hope that the resulting neural activity will generate a useful percept. But the relationship between an electrical pulse and a visual experience is highly complex. Neighboring electrodes can influence one another, individual neurons may respond differently from one moment to the next, and the same stimulation pattern may produce different percepts depending on the brain’s current state. These factors make it difficult to treat electrodes as simple pixels in a camera-like display.
To address this problem, the researchers trained a deep neural network to predict how the participant’s brain would respond to different combinations of stimulation settings. The model analyzed not only which electrodes were activated and how much current they received, but also the participant’s resting brain activity immediately before each stimulation event. This additional information allowed the system to account for changes in neural excitability—the fluctuating readiness of neurons to respond—rather than assuming that the brain would react identically every time.
Once trained, the model was used in reverse. Instead of merely predicting the effect of a stimulation pattern, it searched for the settings most likely to generate a desired pattern of activity in the visual cortex. This approach, known as model-based or closed-loop control, gives the system a way to select electrical inputs according to their predicted neural consequences. It also offers a potential path toward personalized prostheses, in which the device continuously learns how a particular person’s brain responds and adjusts its signals accordingly.
When the AI-designed patterns were tested in the participant, they reproduced targeted patterns of brain activity more accurately while requiring less electrical current than comparison methods. The findings suggest that carefully optimized stimulation may improve control over the neural circuits involved in visual perception without simply increasing the strength or number of electrical pulses. Lower current requirements could be important for future implants because they may reduce unwanted stimulation, limit energy consumption and improve the precision of the resulting percepts.
The researchers also discovered that recorded brain activity was more informative about what the participant perceived than the stimulation settings alone. After some trials, the participant reported whether he had seen a phosphene and, in certain experiments, described its shape, size, brightness and color. The neural recordings provided a better indication of these perceptual outcomes than knowing only which electrodes had been activated. This result highlights a central challenge in neuroprosthetics: the signal sent into the brain does not fully determine the experience that emerges from it. The brain’s response must also be measured and interpreted.
The work does not yet provide natural sight, and the participant’s perceptions remained limited to simple flashes or shapes rather than detailed images. Nevertheless, the study demonstrates a framework that could make future visual cortical prostheses more responsive, reliable and individualized. The researchers envision systems that monitor neural activity, estimate what a user is likely to perceive and modify stimulation in real time. Such technology could ultimately support people who lost vision after strokes, neurodegenerative disease, injury or inherited eye disorders, particularly those whose visual cortex remains capable of responding. For these users, an implant that adapts to the brain rather than forcing the brain to adapt to a fixed electrical recipe could represent a crucial step toward recovering meaningful visual information.
Subject of Research: Artificial intelligence control of visual cortical prostheses and bionic eyes
Article Title: Deep learning-based control of electrically evoked activity in human visual cortex
Web References: Michael Beyeler, UC Santa Barbara; UC Santa Barbara research background; UCSB visual prosthesis research
References: Neuron; “Deep learning-based control of electrically evoked activity in human visual cortex”; article publication date: 7-Aug-2026
Image Credits: Matt Perko, UC Santa Barbara
Keywords
Visual prostheses, bionic eye, artificial intelligence, deep learning, visual cortex, cortical implants, neuroprosthetics, blindness, brain-computer interfaces, biomedical engineering
