Cebra

What is Cebra?

CEBRA is a machine-learning method designed to compress time series data, revealing hidden structures in neural and behavioral variability. It excels at analyzing simultaneously recorded neural and behavioral data, offering a powerful tool for neuroscience research. The method leverages joint behavioral and neural data in a hypothesis- or discovery-driven manner to produce consistent, high-performance latent spaces. CEBRA can be applied to both calcium imaging and electrophysiology datasets, across sensory and motor tasks, and in simple or complex behaviors across species. It supports single and multi-session datasets for hypothesis testing or can be used label-free.

Key features include the ability to decode neural activity from visual cortex to reconstruct viewed videos, decode trajectories from sensorimotor cortex in primates, and decode position during navigation. For example, applying CEBRA to rat hippocampus data yields a median absolute error of 5 cm on a 160 cm track. The method also enables interactive 3D visualizations of embeddings, as demonstrated with mouse primary visual cortex data from the Allen Institute, where 2-photon and Neuropixels recordings are embedded using DINO frame features as labels for video frame decoding.

Benefits include high accuracy in decoding, flexibility across different data types and tasks, and the ability to uncover complex kinematic features and spatial mappings. Use cases range from basic neuroscience research to brain-machine interfaces, where rapid, high-accuracy decoding is essential. Technical details: CEBRA uses a contrastive learning framework with a novel loss function that incorporates behavioral labels or time-based sampling. It is implemented in Python with PyTorch and is available as an open-source package with comprehensive documentation and Colab notebooks for easy experimentation. The method has been validated in Nature 2023 and subsequent studies, demonstrating its utility for mapping space, uncovering kinematic features, and decoding natural movies from visual cortex.

Use Cases

Key Features

time series compression hidden structure discovery joint neural behavioral analysis hypothesis driven latent spaces discovery driven latent spaces calcium imaging support electrophysiology support multi session analysis label free embedding 3D visualization

Who is it for?

neuroscience researchers, computational neuroscientists, behavioral scientists, data scientists, machine learning engineers, neurobiologists

Similar Tools