Deep Learning Steers 96-Electrode Visual Cortex Implant Toward Predictable Phosphenes

A deep learning system has improved the precision of an experimental visual cortical prosthesis in a blind 27-year-old man, generating intended patterns of neural activity more accurately than comparison methods while using less electrical current. The proof-of-concept study, published online in Neuron, involved a 96-electrode array implanted in the visual cortex and tested at Hospital IMED Elche in Spain in 2024.

The work comes from a three-institution collaboration: UC Santa Barbara, ETH Zurich and Miguel Hernández University. Co-first authors are Pehuén Moure (ETH Zurich), Jacob Granley (UC Santa Barbara) and Fabrizio Grani (MHU), supervised by Michael Beyeler, Shih-Chii Liu and Eduardo Fernández. The participant lost his sight after a traumatic brain injury and carried the implant, designed to remain in place for six months, as part of a larger feasibility trial in Spain.

The device could do two things at once: stimulate the visual cortex with electrical current and record the resulting neural activity. That dual capability is the study's practical hinge.

What does this change for cortical prostheses?

Visual cortical prostheses bypass the eyes and optic nerves entirely, delivering electrical signals directly to the visual cortex. That makes them a candidate option for people whose visual pathways are damaged but whose cortex still responds — including patients blinded by strokes, neurodegenerative disease or traumatic brain injury.

The known problem is reliability. The same stimulation does not always evoke the same percept. Activating a given electrode does not guarantee the person will perceive a specific pattern. When researchers stimulated this participant's electrodes, he reported phosphenes — spots, flashes or shapes of light, sometimes resembling stars or fireworks — but predicting them was another matter.

The team trained a deep neural network to predict how the visual cortex would respond to different stimulation settings. The model also incorporated the brain's resting activity immediately before each stimulation, capturing the participant's current neural state. From these predictions, the researchers selected electrode combinations most likely to produce particular activity patterns. The AI-chosen patterns outperformed comparison methods on accuracy and did so at lower current.

One finding deserves a reading-room analogy. Recorded brain activity predicted what the participant perceived better than the stimulation parameters themselves. Input, in other words, is not perception.

"Engineers naturally want to treat phosphenes like pixels: stimulate more electrodes, and you should get a more complete image," Beyeler said. "But the brain does not work that way. Electrodes interact, neural responses fluctuate, and what we put into the brain is not necessarily what the person perceives. The challenge is to learn more about that transformation."

Where the feedback came from

Training used the brain's actual responses, not only electrode identity and current levels. The participant also reported what he experienced — whether he saw a phosphene at all, and in some sessions its shape, size, brightness and color. Comparing those reports with the recordings confirmed that neural signals carried more percept-relevant information than the electrical settings did.

The state-dependence finding may matter most. Brain responses to electrical signals vary over time, including day to day. A model that reads resting activity before each pulse offers a route to stimulation that adapts rather than repeats.

"A useful visual prosthesis cannot rely on a fixed recipe," Beyeler said. "It has to learn how an individual brain responds and adapt the stimulation accordingly. Ultimately, the device should adapt to the person, not the other way around."

The work builds on computational research supported by Beyeler's 2022 NIH Director's New Innovator Award, a five-year, $2 million grant. "Seeing it shape an experiment with a person is something else entirely," he said of moving from abstract models to a human experiment.

The caveats are plain. The technology is investigational, and the participant's percepts were limited to flashes and shapes of light. No FDA-cleared device is involved, and this is a single human subject within a feasibility trial. Still, the demonstration that closed-loop, state-aware AI can improve stimulation precision marks a concrete step toward prostheses that learn each brain rather than assume one. The next question is whether that approach scales beyond one cortex and six months.