The Variance Brain Foundation Models Forgot: Third-Order Statistics Predict Cognition Where Billion-Parameter Models Fail
arXiv:2606.04010v1 Announce Type: cross Abstract: Brain foundation models (BFMs) are self-supervised Transformers pretrained on fMRI data. We posit that these models should capture each subject's cognitive performance from their fMRI signal. Yet across three state-of-the-art BFMs and every readout we test, they predict cognition worse than a linear regression from the $\sim$80K parameters of the functional connectivity matrix (FC). The gap widens with scale: BrainLM's 650M model predicts cognition worse than its 111M. We attribute this to a \textbf{variance allocation problem}: BFM pretraining
This research highlights a fundamental limitation in current large AI models applied to complex biological data, particularly brain imaging, suggesting a potential architectural blind spot that only now is becoming evident through extensive experimentation.
It suggests that simply scaling up parameters in current AI architectures may not be sufficient for understanding complex biological systems like the human brain, indicating a need for new theoretical foundations and model designs for AI in biology.
The prevailing assumption that more parameters automatically lead to better performance is challenged, particularly in nuanced domains like brain-computer interfaces or medical diagnostics, forcing a re-evaluation of current BFM development strategies.
- · Neuroscience researchers focused on statistical methods beyond mean-field
- · Developers of novel AI architectures for biological data
- · Small data/sparse data AI specialists
- · Companies heavily invested in scaling existing BFM architectures
- · Purely 'brute-force' scaling approaches to AI
- · Developers neglecting biological nuances in AI design
Brain Foundation Models (BFMs) are found to perform worse than simpler statistical methods in predicting cognition, challenging their efficacy for real-world neurological applications.
This could lead to a strategic pivot in AI research for biological data, emphasizing more biologically informed architectures and statistical methods beyond high-dimensional mean-field approaches.
Future brain-computer interfaces and neuro-AI applications might adopt hybrid models combining deep learning with advanced statistical mechanics, potentially accelerating development by avoiding scale-only dead ends.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI