Research

Foundation models for biosignals

Every EEG dataset uses a different electrode montage, and every wearable has a different channel count. I build pre-trained models that read any layout into a shared latent space, so one model serves many devices and tasks with minimal fine-tuning.

LUNA introduced query-based channel unification at NeurIPS 2025. FEMBA swapped quadratic attention for linear-time Mamba. LuMamba combines both and adds LeJEPA world-model pre-training. TinyMyo extends the family to EMG. All are open source in BioFoundation with weights on Hugging Face.

Key numbers

  • 300× fewer FLOPs than standard transformers (LUNA)
  • 21,000+ h of EEG in pre-training
  • 0.97 AUPR Alzheimer's detection on unseen montages (LuMamba)

Tiny recursion models

Reasoning-style computation does not have to mean billions of parameters. I study how deep supervision and adaptive-depth recursion let very small networks refine their answers iteratively, spending compute only where the signal demands it.

Recent work reframed TRM recursion as annealed sampling on an energy-based model, running the loop on a thermodynamic-computing simulator. Two open MSc topics extend TRMs to time series and quantized edge deployment.

Key numbers

  • fewer sampling sweeps via recursion (Thermo-TRM)
  • 2nd place ETH Probabilistic Computing Hackathon

Ultra-low-power deployment

A model is only clinically useful if it runs where the patient is. I co-design algorithms and implementations for GAP9 and RISC-V platforms: quantization, on-device continual learning, and energy-aware architectures for seizure detection, BMI, and speech imagery.

This line of work goes back to my PhD (EEG-TCNet, EpiDeNet, BioGAP) and continues with quantized foundation models running on wearables.

Key numbers

  • µW-scale inference on GAP9
  • SOTA wearable seizure detection benchmarks
  • BCI Award nomination for drone-control headband