#20 · ML / AI methods

Learned latent representations of brain recordings are evaluated beyond predictive fit

The dominant thread is the use of learned latent representations of neural recordings for individual-level inference. Several items build these representations directly: BRAID-fMRI aligns fMRI responses across 8 datasets and participants into a shared CLIP space using a CLIP-supervised ROI-wise Transformer, while NeuroCBIR uses a VAE plus contrastive learning to retrieve whole-brain and region-specific 3D T1w MRI. MovieSTAGE and CoHyFuse encode movie-fMRI and task-fMRI functional connectivity as hypergraphs, with whole-movie and whole-session FC as global context, for ADHD-related classification and prediction. The EEG item extends this thread by evaluating latent structure beyond predictive fit: per-subject sparse dynamics on EEG foundation-model latents show denser epilepsy graphs despite matched predictive fit (n=198), indicating that matched performance does not establish equivalent latent dynamical structure.