Neural Noise

AI-driven bidirectional visual neuroprosthesis

2 min read

  • An AI-driven brain implant brings “two-way” artificial vision (read and write) closer to reality. 1
  • AI is integral to encoding and decoding in this bidirectional visual neuroprosthesis. 1
  • The approach is relevant to sensory neuroprosthetics and cortical interfaces. 1 1

Weekly enrichment (2026-07-20)

  • The underlying work is a Science Advances study from Miguel Hernández University (UMH, Eduardo Fernández Jover’s group) reporting phosphene-perception results in blind volunteers as a step toward a bidirectional cortical visual prosthesis.2 3
  • The interface uses an intracortical Utah microelectrode array of 100 electrodes (96 recording channels) implanted in occipital/early visual cortex, letting the same array both stimulate and record nearby neural activity.2 4
  • It is closed-loop rather than open-loop: while delivering electrical patterns that evoke visual sensations (“writing”), it simultaneously records neuronal responses (“reading”) and adapts stimulation to nearby neurons’ reactions in real time.3 5
  • A key result is that subjective visual experience could be accurately predicted from recorded neural activity, so the system can infer whether a given stimulation will produce a phosphene and estimate its brightness and the number of percepts.2 5
  • With this bidirectional exchange, implanted participants recognized complex patterns, movements, shapes, and even some letters, and the feedback loop accelerated the users’ learning curve.3 5
  • The companion AI/deep-learning method (bioRxiv, 2025) trained a ~1.2M-parameter forward neural network on a 96-channel Utah array in a blind participant to predict single-trial evoked responses (ΔMUAe) from stimulation parameters.4
  • That dataset spanned 5,818 random and 484 structured multi-electrode stimulation patterns delivered across 26 days over four months, with per-day resting-state activity included to handle day-to-day drift.4
  • Two control strategies—a gradient-based optimizer and a learned inverse network for real-time stimulation synthesis—both beat 1-to-1 and linear baselines (forward model p < 0.0001), elicited more consistent percepts, and required lower stimulation currents.4
  • The forward model outperformed even an idealized “optimal dictionary,” strong evidence that multi-electrode phosphenes are not the linear sum of single-electrode responses, reflecting nonlinear cortical integration.4
  • Stated clinical goal is functional vision (navigation, mobility, reading large characters) rather than natural sight, with the read-and-adapt approach aimed at safer, more stable, scalable cortical vision restoration.3 5

Footnotes

  1. https://news.google.com/rss/articles/CBMib0FVX3lxTE9iSVVnNlRDMk1HRzZ3dHlPSGZkcktCempEVXVQLWtDRURMUDM5bVFiY1RnVlNhUWg1UE1PUVhMdk5pcTlsUUkydzZfOU9CbzlWWFh6aG4zS2xweVZGUFk2NXJaOUFTcm9tQV9hRU5ka9IBdEFVX3lxTE5YQ25TUDBMeTc4YjZlVndsaXBVNlMzcXpXRmNVcnZqMmgzZkhlMmx6NmVnVVhKMnIxQVhxamVFQmdGNGdURFo4aDFscmZyMlVTY21LVnJLVmlCNTVJTER4RUtvUGluRVU2Q25sMmVLc0VpaWNT?oc=5 ↩ ↩2 ↩3 ↩4

  2. https://doi.org/10.1126/sciadv.adv8846 ↩ ↩2 ↩3

  3. https://neurosciencenews.com/ai-neuroprosthetics-vision-29904/ ↩ ↩2 ↩3 ↩4

  4. https://www.biorxiv.org/content/10.1101/2025.09.24.678361v2 ↩ ↩2 ↩3 ↩4 ↩5

  5. https://www.eurekalert.org/news-releases/1104900 ↩ ↩2 ↩3 ↩4

Graph View