AI decodes brain activity
A new AI technology from EPFL provides deeper insights into neuronal dynamics. Geometric deep learning enables scientists to decode universal patterns of brain activity for the first time. This has far-reaching applications in neuroscience, robotics and other disciplines.
Scientists at EPFL have developed a groundbreaking AI technique based on geometric principles that visualises neuronal activity patterns. The system, called MARBLE (Manifold Representation Basis Learning), breaks down electrical signals from the brain into dynamic patterns that can be decoded by a neural network. This makes it possible to recognise universal strategies of brain activity across different individuals and experimental conditions.
Visualising hidden patterns in the brain
Neuroscience is facing a fundamental challenge. Brain activity is usually recorded by analysing a few neurons, which means that a complete picture of neuronal processes is lacking. Pierre Vandergheynst, head of the LTS2 signal processing laboratory at EPFL, compares this problem to the story of blind people feeling different parts of an elephant and drawing contradictory conclusions. The situation is similar with the recording of neuronal signals; a limited data section makes overall understanding more difficult.
The system has now been able to show that different animals that used the same mental strategies to solve problems exhibited matching neuronal patterns. The technique thus enables a more precise interpretation of brain activity and could set a new standard for analysing dynamic neuronal processes.
A breakthrough for neuroscience and robotics
The innovative approach of geometric deep learning makes it possible to analyse neuronal data not only statistically, but also in its natural mathematical context. This shows that brain activity can be visualised as complex geometric structures. For example, in the form of a torus, similar to a donut.
The EPFL researchers tested MARBLE with recordings from the macaque premotor cortex during grasping movements and in the hippocampus of rats during spatial orientation tasks. The results were impressive. The system decoded the neuronal activity far more precisely than conventional methods and enabled a more intuitive interpretation of the neuronal processes.
Broad application potential beyond neuroscience
In addition to its use in brain research, MARBLE could also be of great value to other scientific disciplines. The technology offers the possibility of converting neuronal activity patterns into decodable signals. This can be used to control robotic assistance systems that react to brain activity.
Pierre Vandergheynst emphasises the potential beyond neuroscience: “Our method is based on the mathematical theory of high-dimensional structures and can also be used in other scientific disciplines to analyse dynamic processes and identify universal patterns.”
MARBLE could represent a fundamental step forward in the study of complex biological and physical systems, not only revolutionising our understanding of the brain, but also providing new impetus for artificial intelligence and robotics.