Tip Of AI
Discover what you can do with AI
CEBRA (Learnable latent embeddings for joint behavioural and neural analysis) is a machine-learning method that offers a novel approach to compress time series data, shedding light on hidden structures in the variability of the data. It specializes in analyzing and decoding behavioral and neural data recorded simultaneously, allowing researchers to reconstruct viewed videos from the activity in the visual cortex of the mouse brain. One of the key features of CEBRA is its ability to produce consistent and high-performance latent spaces by jointly utilizing both behavioral and neural data. It can be applied in a supervised or self-supervised manner, either driven by specific hypotheses or aimed at discovery. The method proves effective in decoding neural dynamics by uncovering meaningful differences and inferring latents from the data. CEBRA has been validated and shown its accuracy across various datasets, including calcium and electrophysiology data, as well as sensory and motor tasks in both simple and complex behaviors across different species. It can be used with single or multi-session datasets for hypothesis testing or label-free analysis. Moreover, CEBRA demonstrates its versatility in mapping space and revealing complex kinematic features. It exhibits consistency in latent spaces across 2-photon and Neuropixels data and achieves rapid and high-accuracy decoding of natural movies from the visual cortex. The pre-print of the paper detailing the CEBRA algorithm is available on arxiv.org/abs/2204.00673, and the official implementation can be found on GitHub.