A machine learning algorithm called CEBRA is unlocking the hidden structure in neural code, with the potential to revolutionize brain-machine interfaces (BMIs). The algorithm can be used to decode what a mouse sees while watching a movie, predict primate arm movements, and reconstruct the positions of rats as they roam. The research, led by Mackenzie Mathis, EPFL’s Bertarelli Chair of Integrative Neuroscience, was published in Nature.
CEBRA learns the latent structure of a mouse’s visual system by predicting unseen movie frames from brain signals alone, after an initial training period. The data, provided by the Allen Institute in Seattle, comes from either direct brain activity measurements or optical probes using genetically modified mice with activated neurons that glow green. Impressively, CEBRA performs well with less than 1% of neurons in the visual cortex.
The algorithm is based on contrastive learning, which arranges high-dimensional data into a lower-dimensional latent space. This enables researchers to infer hidden relationships and structures, such as neural data and behavioral labels. CEBRA excels at reconstructing synthetic data, combining data across modalities, and limiting nuances. Steffen Schneider, co-first author of the paper, explains that the algorithm’s strength lies in its ability to combine movie features and brain data.
CEBRA’s primary goal is to uncover structures in complex systems like the brain, offering insights into information processing and new principles in neuroscience. However, its applications extend beyond neuroscience; it can be applied to datasets involving time or joint information, including animal behavior and gene-expression data. The potential clinical applications of this versatile algorithm are truly exciting.
What’s next? As AI improves (which it is doing at an exponential pace), we will see technologies like CEBRA evolve into unimaginably powerful pattern-matching tools. What we will (or may) do with them is easy to imagine.
If you want a foundational understanding of machine learning and generative AI, enroll in our free online course, Generative AI for Execs. It will help you understand the power of pattern-matching algorithms and how you can use them to increase productivity and create value for your business.