Foundation Models

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 our full-stack framework for deploying biosignal foundation models on ultra-low-power edge devices at AMLD 2026 in Lausanne.

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

Presented quantized biosignal foundation models in EMBC’s silent disco session while Anna Tegon delivered FEMBA.

Transfers EEG foundation models to ECG/PPG for BP estimation, reaching 1.57 mmHg MAE with 3.5× model compression via INT8 quantization.

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