Cebra

CEBRA is an open-source Python library for compressing high-dimensional time-series data, such as neural recordings and behavioral videos, into low-dimensional, interpretable latent spaces. Designed for neuroscientists, computational biologists, and ML researchers, it excels on data where behavior and neural activity are recorded simultaneously. The method is grounded in a Nature 2023 paper and has been validated for decoding viewed videos from the visual cortex, reconstructing trajectories from primate sensorimotor cortices, and mapping position during navigation. CEBRA operates in two modes: supervised (hypothesis-driven) using auxiliary behavioral labels, and self-supervised (discovery-driven) for label-free analysis. It supports hybrid embeddings, consistency metrics for comparing latent spaces, and k-nearest neighbor decoding. The library handles calcium imaging, electrophysiology (including Neuropixels), and behavioral data across species, with demonstrated accuracy—e.g., median absolute error of 5 cm on a 160 cm track length in rat hippocampus data. Key features include a scikit-learn-compatible API for easy integration into Python pipelines, built-in plotting with matplotlib and plotly, GPU-accelerated training via PyTorch, and a Docker container for reproducibility. The latest extension (AISTATS 2025) adds time-series attribution maps using regularized contrastive learning, enabling interpretability of which time points drive the embedding. Compared to generic embedding methods like UMAP or t-SNE, CEBRA explicitly incorporates behavioral or auxiliary variables as labels, producing consistent, interpretable embeddings optimized for neural decoding and hypothesis testing. This makes it a research-grade tool for those working with neural-behavioral datasets, while its open-source nature and active development (GitHub) foster community adoption.