Cebra

CEBRA is a self supervised machine learning method for producing interpretable, consistent embeddings from high dimensional recordings, with a focus on neuroscience. It takes time series data, such as neural activity recorded alongside behavior, and compresses it into a lower dimensional latent space that reveals structure otherwise hidden in the variability of the data. The method is described in a 2023 Nature paper by Schneider, Lee and Mathis. The site is aimed at researchers who analyze neural and behavioral data. Demonstrations on the page include decoding position from rat hippocampus recordings, reconstructing a viewed video from mouse visual cortex activity, and decoding trajectories from primate sensorimotor cortex. A later extension adds attribution maps for time series data using regularized contrastive learning. The page links to documentation and demos, the source code, and interactive notebooks that open in Colaboratory. The vendor does not publish prices. No pricing page or plan information is available on the site. The page does carry a patent notice stating that EPFL has filed a patent application on dimensionality reduction of time series data and the resulting embeddings, and it directs readers with non academic use in mind to take note of that notice before relying on the method.