EEG

Lightning talk on LuMamba — using LeJEPA “world-model” self-supervision to pretrain a tiny, montage-agnostic foundation model for biosignals.

We bring LeJEPA to biosignals for the first time and show that the choice of self-supervised objective (masked reconstruction, LeJEPA, or a mix) trades latent-space structure for cross-montage generalisation, all on a 4.6M-parameter bi-Mamba model that is 377× cheaper than LaBraM.

How learned queries and cross-attention let LUNA decouple compute from electrode count, cutting FLOPs by 300× and GPU memory by 10× while still hitting state-of-the-art on TUAR.

Poster presentation of LUNA—our topology-agnostic EEG foundation model—at NeurIPS 2025 in San Diego.

Topology-agnostic EEG foundation model that delivers state-of-the-art performance with 300× fewer FLOPs and 10× lower memory usage.

Bidirectional Mamba EEG foundation model that matches transformer accuracy while scaling linearly with sequence length.

Community-driven framework standardizing datasets, metrics, and tooling for reproducible EEG seizure detection research.

Alternating-attention EEG foundation model pre-trained on 20k+ hours that doubles speed and cuts memory by 6× versus standard transformers.

Open-source framework powering LUNA, FEMBA, CEReBrO, and other EEG foundation models — including LUNA, accepted at NeurIPS 2025.

BERT-inspired EEG encoder that cuts seizure-detection false positives to 0.23 FP/h after subject-specific tuning.