Meta's Fundamental AI Research (FAIR) lab has released NeuralSet, a Python package designed to bridge the gap between neuroscience data and deep learning workflows. The framework addresses a major pain point in Neuro-AI research: the cumbersome process of preparing brain data—such as fMRI, M/EEG, and spike recordings—for use with modern AI models.
Traditional neuroscience tools like MNE-Python and Nilearn are optimized for signal processing but lack native support for deep learning pipelines. Researchers often resort to building ad-hoc systems to align neural time series with embeddings from AI frameworks like HuggingFace Transformers, leading to inefficiencies as datasets grow to terabyte scale.
NeuralSet introduces a structure–data decoupling approach. Its core abstractions—Events, Extractors, Segments, Batch Data, and a Backend layer—allow users to work with lightweight metadata before loading raw signals. Events capture experimental components (e.g., an fMRI run or a spoken word) with attributes like type, start time, duration, and timeline. A Study object organizes all events into a pandas DataFrame, enabling efficient filtering and exploration without memory overhead.
The package supports BIDS-compliant datasets and provides composable EventsTransform operations for tasks such as cross-validation splitting or chunking long stimuli. NeuralSet is open-source and available on GitHub.