Scientists at the Swiss research institution EPFL have developed a novel machine-learning algorithm, called CEBRA, capable of predicting what mice see with remarkable accuracy by decoding their neural activity. Furthermore, the algorithm can be used to predict movements of the arm in primates and reconstruct the positions of rats in an arena.
CEBRA works by mapping brain activity to specific frames, and can predict unseen movie frames directly from brain signals alone after an initial training period. The algorithm is based on contrastive learning, a technique that learns how high-dimensional data can be arranged or embedded into a lower-dimensional space, called a latent space, which allows for similar data points to be close together and more distinctive data points to be further apart.
To demonstrate the performance of their algorithm, the EPFL researchers tested CEBRA’s accuracy by decoding what a mouse saw while watching a movie. Using raw neural data, the team found that they were able to infer hidden relationships and structure with CEBRA, and were able to accurately reconstruct movies with only 0.5 million neurons in the mouse’s visual cortex.
In addition to its success in decoding visual information, researchers applied CEBRA to predicting movements of the arm in primates and reconstructing the positions of rats as they moved around an arena – showing potential clinical applications with the novel algorithm.
The EPFL team, led by Mackenzie Mathis, PI of the study and Bertarelli Chair in integrative neuroscience, published their work in Nature.
Mackenzie Mathis expressed enthusiasm for the implications of CEBRA, commenting that “CEBRA excels compared to other algorithms at reconstructing synthetic data, which is critical to compare algorithms. Its strengths also lie in its ability to combine data across modalities, such as movie features and brain data, and it helps limit nuances, such as changes to the data that depend on how they were collected.”
The team’s study, entitled “Learnable latent embeddings for joint behavioural and neural analysis” highlights CEBRA’s potential in uncovering the hidden structure and relationships in complex systems, particularly the brain. As the algorithm is not limited to neuroscience research and can be applied to many datasets, its potential clinical applications are vast.
