CONFORM: A Project to Create Crowd-Sourced Open Neuroscience fMRI Foundation Models

Michael J. Tarr, Jacob S. Prince, Ben Lahner, Mayukh Deb, Aude Oliva, Apurva Ratan Murty, John A. Pyles, Margaret Marie Henderson, Leila Wehbe, Andrew Luo

Keywords: fMRI, Foundation Models, generalization, preprocessing, dataset

TL;DR: CONFORM is a crowd-sourced framework to build a human fMRI foundation model through improved preprocessing, dataset aggregation, and in-context representation learning.

Abstract: We propose CONFORM (Crowd-Sourced Open Neuroscience fMRI Foundation Model), a project that will bring together recent advances in neural data processing and analysis with a novel, crowd-sourced infrastructure. This transformative approach will overcome several current challenges in creating a foundational human fMRI model for vision: collecting massive amounts of data from a handful of participants is neither scalable nor sustainable; the number of participants is small for such datasets; stimulus diversity is limited; and generalizability to different populations is poor. CONFORM will overcome these limitations by combining a powerful generative denoising method (SNAP), a scalable framework for aggregating existing fMRI datasets (MOSAIC), and a meta-learning model that enables generalization with much smaller data from new participants (BraInCoRL). Our collaborative effort will produce models built on unprecedented scale and diversity—ultimately with hundreds of participants and hundreds of thousands of naturalistic image and movie stimuli—and provide the tools for continuous expansion of the underlying dataset. This ``crowd-sourced'' approach will allow many more researchers to leverage state-of-the-art NeuroAI methods using the scale of data they typically collect, democratizing access to powerful models and accelerating scientific discovery for a wide range of neuroscientific domains and populations.

Submission Number: 28