TorchBrain

Overview. TorchBrain (https://torchbrain.org/) is an open-source python package designed to make the modern deep learning toolbox and neuro-foundation models accessible for both neuroscientists and ML practitioners. This tool efficiently scales model training to large and diverse datasets (TB-scale data), makes neural data easier to use by providing access to ready-for-training datasets across multiple brain modalities, and unlocks new frontiers in neurotech. We believe that providing researchers with this user-friendly, robust and reliable engineering layer is critical for building and adopting foundation models.

Why this matters. Progress in language modeling was in large part made possible thanks to major engineering efforts put into making high performance training and inference pipelines accessible, and we are replicating that in neuroscience. Our goal is to lower the barriers to using powerful models, standardize access to datasets, and accelerate research. Already, we have seen the tools we’ve developed catalyze research across multiple groups.

The demo. Our goal is to show attendees how these tools make cutting-edge methods approachable and usable in real-world research workflows. Specifically, the demo will highlight how researchers can: 1) Rapidly fine-tune pre-trained models on their own datasets. 2) Leverage standardized access to diverse datasets covering a wide range of neural modalities (EPhys, OPhys, EEG, iEEG). 3) Explore the use of neural embeddings to gain scientific insights.

Concretely, we will walk attendees through an interactive demo demonstrating how users can leverage TorchBrain to manage multiple streams of data and deploy a variety of samplers to train their deep learning models. Our booth will also act as a help desk, where attendees can share their use cases, and we help them discover how they can use our open-source tools in their own workflows to streamline data access, train and fine-tune models, and ultimately accelerate neurotech and scientific discovery in neuroscience.