EPFL researchers are advancing the development of AI-driven vision prosthetics capable of restoring meaningful, object-level sight to the blind. Led by Martin Schrimpf at the NeuroAI Lab, the team utilizes artificial intelligence to predict precise brain stimulation patterns required to evoke images of faces and specific objects, moving beyond the limited light flashes produced by current technologies. Traditional visual prosthetics, such as retinal or optic nerve implants, often fail when damage occurs in the retina or optic nerve, necessitating cortical implants that bypass these structures entirely. Current cortical devices stimulate lower-level brain regions, resulting only in simple shapes or light spots due to hardware constraints and the complexity of targeting higher visual areas. These higher regions are essential for processing complex objects like cars or houses but are difficult to access because their specific stimulation requirements remain unknown. The EPFL team addressed this challenge by developing topographic neural networks. These models simulate how different stimulation patterns in higher-level brain regions influence perception, allowing researchers to identify the optimal electrical signals needed to evoke specific images without extensive, costly physical trials. Johannes Mehrer, the lead scientist, explained that while existing methods cannot generate complex object percepts, their approach allows for the manipulation and selection of image representations within the brain. The preliminary results of this research were presented in April at the International Conference on Learning Representations in Rio de Janeiro. Following the computer simulations, Dutch researchers in Amsterdam validated the model by testing it on two sighted monkeys that already had brain implants for other experiments. The study demonstrated that the AI model efficiently predicted stimulation patterns that significantly influenced the monkeys' visual object recognition. The researchers successfully shaped object perception, effectively distorting how the monkeys interpreted visual stimuli. However, the team emphasizes that they have not yet achieved the ability to create object perception from scratch. The current experiments involved a monkey viewing an existing image, which the researchers then altered through stimulation. The next critical step is to evoke a meaningful percept without any external visual input, thereby restoring sight to individuals whose eyes cannot deliver usable images. Martin Schrimpf noted that this modeling technique could also revolutionize hearing prosthetics. With funding from the Horton Health Foundation, the team plans to investigate whether topographic models can similarly predict neural activity for auditory processing, aiming to improve upon current cochlear implants that do not fully restore natural hearing capabilities. While full human trials are a future goal, these findings mark a significant leap toward prosthetics that offer more than just basic light perception, potentially enabling blind individuals to recognize faces and objects. The full research paper is available on the arXiv preprint server.