Advancements in AI Models for Visual Prosthetics
This research from the NeuroAI Lab at EPFL, under the leadership of Martin Schrimpf, explores the use of artificial intelligence to enhance visual prosthetics. The study focuses on predicting precise brain stimulation sites to evoke images of faces and specific objects, moving beyond mere flashes of light. Dutch researchers conducted live trials using the models developed at EPFL, yielding promising preliminary results presented at the International Conference on Learning Representations in April. These findings suggest significant implications for improving human vision.
Addressing Visual Impairments
Johannes Mehrer, a scientist at the NeuroAI Lab and the research leader, articulated the motivation driving this project. Many individuals experience irreversible visual deficits along the visual processing pathways, starting from the retina. To address this issue, developing visual prosthetics may offer a viable solution. These devices vary in type, including retinal, optical nerve, and cortical prosthetics, tailored to the extent of the damage within the visual system.
The Types and Limitations of Visual Prosthetics
Retinal prosthetics are intended for placement on the retina, while optical nerve prosthetics serve as alternatives when the retina is too damaged. In cases where neither can be utilized, cortical prosthetics stimulate the visual cortex directly, bypassing earlier stages of processing. Unfortunately, existing prosthetic designs are constrained; they primarily project simple symbols and light flashes onto lower-level brain regions. Consequently, they are unable to evoke the perception of complex visual objects like houses or cars.
Exploring Higher-Level Visual Processing
Higher-level visual regions in the brain facilitate the processing of intricate objects, making them ideal targets for next-generation prosthetics. However, access to these areas is challenging due to the lack of knowledge regarding optimal stimulation techniques. The AI model developed by the EPFL researchers aims to overcome this barrier by identifying effective stimulation patterns for these advanced regions.
Innovative Approaches Using Neural Networks
The research team employed a specialized topographic neural network to simulate various brain stimulation patterns in higher-order visual areas. This approach allows them to efficiently test numerous combinations of parameters without extensive experimental costs or time. Based on the model’s predictions, researchers in Amsterdam applied these insights during trials with sighted monkeys that already had implants from unrelated studies.
Shaping Object Perception with AI
Martin Schrimpf, head of the NeuroAI Lab, noted the model’s success in predicting effective stimulation patterns that influenced the monkeys’ visual object recognition. While the researchers have demonstrated the ability to shape object perception—altering existing visual stimuli—the next goal is to evoke meaningful perceptions where none currently exist. This would represent a significant breakthrough in restoring vision for individuals who are blind.
Future Applications and Challenges
This groundbreaking research could also extend beyond visual prosthetics to enhance auditory processes. With continued support from the Horton Health Foundation, Schrimpf and his team are evaluating whether similar AI modeling techniques can optimize auditory stimulation. While cochlear implants have made strides in auditory restoration, they are not flawless, prompting the exploration of topographic models to refine neural responses for improved auditory processing.
