Companion preprint to the ongoing doctoral work on EEG microstate analysis with deep generative models. The paper replaces the classical k-means clustering of microstate topographies with a Convolutional VaDE that jointly performs topographic reconstruction and probabilistic soft clustering, evaluated through a systematic sweep over architecture and cluster-count hyperparameters.
← Publications · 2026 · preprint
Interpretable EEG Microstate Discovery via Variational Deep Embedding: A Systematic Architecture Search with Multi-Quadrant Evaluation
Saheed Faremi · Andrea Visentin · Luca Longo
arXiv
Abstract
A Convolutional Variational Deep Embedding (Conv-VaDE) model for analysing brain electrical activity. The approach jointly learns topographic reconstruction and probabilistic soft clustering in a shared latent space, replacing traditional hard-assignment methods with probabilistic alternatives. Through a systematic architecture search across cluster counts (K=3-20), latent dimensionality, network depth, and channel width, optimal results are reported at depth L=4, achieving a best-case GEV of 0.730 and a silhouette of 0.229 at K=4. Tested on the LEMON resting-state EEG dataset.