Rethinking Brain Models: Linear Simplicity Wins Over Complex Transformers

The quest to understand the human brain through artificial intelligence has hit another snag. Brain foundation models (BFMs), with their self-supervised Transformer architecture, are struggling to predict cognitive performance from fMRI data. Despite their sophisticated design, these models are outperformed by a more straightforward linear regression approach using a functional connectivity matrix (FC) of roughly 80,000 parameters.

Complexity Isn't Always Better

Three leading BFMs have shown that increasing model size doesn't equate to better performance. The BrainLM model, with its hefty 650 million parameters, actually predicts cognition worse than its more modest 111 million parameter version. So what's going wrong here?

The issue seems to be a 'variance allocation problem'. The BFMs are capturing the variance components that flood the fMRI data, but miss out on the higher-order structures essential for predicting cognition. It's like trying to hear a whisper amidst a cacophony, the models are too busy processing the noise to catch the important signals.

The Linear Edge

Enter the simple solution: a linear pipeline that effectively projects the fMRI signal into a subspace preserving its co-skewness. This approach doesn't require pretraining or GPUs, yet it outshines every BFM on all tested datasets and parcellations. It's a stark reminder that enterprise AI is boring. That's why it works.

In fact, by fine-tuning BrainLM's forward pass using a targeted loss aimed at this subspace, the ceiling of raw FC predictions can be reached. This indicates the primary bottleneck isn't the architecture or model size, but the pretraining objectives themselves. When will we stop overestimating the complexity needed for effective AI?

The Path Forward

What does this mean for those developing AI in cognitive neuroscience? It's simple: re-evaluate the necessity of model complexity. Trade finance is a $5 trillion market running on fax machines and PDF attachments, and sometimes, the simplest methods prove to be the most effective. In the race to decode the brain, perhaps a recalibration towards simplicity could be key.

So, the next time you're tempted to bet on bigger and more complex models, ask yourself: is the ROI in the model, or in the 40% reduction in document processing time? The brain may be complex, but our approach to understanding it doesn't have to be.