NeuralSet: A High-Performing Python Package for Neuro-AI

Neuroscience is increasingly leveraging artificial intelligence to decode brain activity, but researchers face a fragmented software ecosystem where tools are siloed by recording modality (e.g., fMRI, M/EEG, spikes) and optimized for small, in-memory datasets. This bottleneck prevents the use of massive, naturalistic data and hinders integration with modern deep learning. A new paper from Jean-Rémi King and a team of 28 researchers introduces NeuralSet, a Python framework designed to unify diverse neural recordings and experimental stimuli into a single, scalable pipeline.

NeuralSet decouples experimental metadata from lazy, memory-efficient data extraction, enabling seamless handling of datasets that exceed RAM limits. It harmonizes standard neuroscientific preprocessing workflows with pretrained deep learning embeddings, supporting inputs ranging from text and audio to video. The framework provides a unified PyTorch-ready interface, allowing researchers to build and train models without manual data wrangling or modality-specific tools. This design ensures full computational provenance, meaning every preprocessing step is tracked and reproducible.

The package is built for scalability: it works on a single laptop for prototyping and seamlessly scales to high-performance computing clusters for production-level analysis. By eliminating the need to juggle multiple libraries and custom scripts, NeuralSet reduces the barrier to entry for neuro-AI research. The paper is available on arXiv (2605.03169) and includes code, data, and media links. This unified infrastructure promises to accelerate discoveries at the intersection of neuroscience and AI, enabling studies that combine fMRI, EEG, and spike recordings with rich naturalistic stimuli.