The Variance Brain Foundation Models Forgot: Third-Order Statistics Predict Cognition Where Billion-Parameter Models Fail
A new research paper proposes that current brain foundation models (BFMs) fail to capture crucial third-order statistical properties of brain activity, which are vital for predicting cognitive performance. These large-scale models, trained on fMRI data, prioritize capturing dominant variance components over higher-order structures like co-skewness. The researchers developed a linear pipeline that preserves these essential co-skewness tensors, outperforming existing BFMs and raw functional connectivity matrices without requiring extensive pretraining or computational resources. AI
IMPACT This research suggests a new direction for developing more effective brain foundation models by focusing on higher-order statistical properties.