Draft. This entry is pending review.
EEG microstates are quasi-stable scalp topographies (typically four to seven canonical classes) that segment continuous EEG into a discrete temporal alphabet. This work asks whether a learned latent geometry, via a variational autoencoder, produces a microstate segmentation that is more interpretable, more stable across sessions, or more predictive of behaviour than classical clustering.
Variants explored:
- VAE. Single Gaussian latent prior; learns continuous embedding.
- GMM-VAE. Gaussian-mixture latent prior; one component per microstate class.
- Architecture-search experiments to compare codebook capacity, regularisation, and decoder choices.