Since august 2025 I've been with a small independent research team (myself and two collaborators) building a machine learning architecture for EEG-based paradigms.
The diagram below is a direct read of the code path actually instantiated for training
— SparseEvidenceGNNCore's forward pass under the dense_edge_gru
config: a dilated Conv1d channel encoder running alongside a coherence/event lane (cross-wavelet
transform → cone-of-influence masking and surrogate significance → smoothing/downsampling
→ a per-edge GRU), which merge and pass through a message network, destination-node
aggregation, and a final classifier.
A real coherence graph from one preictal window (chb01, seizure 16 — the fold above, auc_pr 0.956, false alarms/hour 1.2), rendered as an interactive 3D scene — drag to rotate, click any highlighted edge for the full coherence / significance / phase detail behind it. Moved to its own page: see Visual.
I keep detailed session notes as I work through this. A few are written up on the Notes page.
This project builds directly on wavelet coherence methods I developed in earlier work, including a peer-reviewed real-time wavelet coherence architecture (IEEE ISPA 2025) and my M.S. thesis on wavelet-based time-frequency fingerprinting — see Publications for both.