Thorir Mar Ingolfsson

Making AI run on microwatts

Thorir Mar Ingolfsson · Postdoctoral Researcher · ETH Zürich

A kilogram and a half of matter, twenty watts of power: the human brain is still the most efficient intelligence in existence. Frontier AI needs a power plant. I work on closing that gap from the bottom up, for the signals the brain itself produces: foundation models for EEG, ECG, and other biosignals, compressed until they run in real time on wearable devices, not in the cloud.

Explore my research Work with me → Download CV
News
Jul 2026VitalFM funded by the Swiss AI Initiative (Large Grant, 400k GPU hours on Alps): a unified foundation model for physiological waveforms, from Alps to edge.
Jul 2026Awarded a Hasler Foundation grant (CHF 50,000) for quantized recursive inference on resource-constrained edge AI.
Jun 20262nd place at the ETH Zürich Probabilistic Computing Hackathon with Thermo-TRM.
Jun 2026LuMamba accepted at EUSIPCO 2026 — brings LeJEPA to biosignals, 377× cheaper than LaBraM.
Sep 2025LUNA accepted at NeurIPS 2025 — topology-agnostic EEG foundation model, 300× fewer FLOPs.
1095 Citations Google Scholar
14 h-index Core research impact
30 Publications View publications →

Three questions, one goal

How do we make machine learning clinically useful on devices that run for weeks on a coin cell? My work attacks this from three directions.

Foundation models for biosignals

Pre-trained models that understand EEG and other biosignals across devices, montages, and tasks, with minimal fine-tuning.

LUNA · NeurIPS 2025 · 300× fewer FLOPs, SOTA on TUAR, any electrode layout.
Explore this direction →

Tiny recursion models

Deep supervision and adaptive-depth recursion for time series: sophisticated temporal reasoning inside microcontroller budgets.

Thermo-TRM · recursion as annealed sampling; 2nd place, ETH Probabilistic Computing Hackathon.
Explore this direction →

Ultra-low-power deployment

Hardware and algorithm co-design that puts clinical-grade models on GAP9 and RISC-V wearables at microwatt power.

Seizure detection & BMI · real-time, on-device, battery-week lifetimes.
Explore this direction →
Flagship software

BioFoundation

★ 121 stars⑂ 15 forksApache 2.0

The open-source home of our biosignal foundation-model family. PyTorch Lightning, Hydra configs, pre-trained weights on Hugging Face, and distributed training: everything needed to pre-train, fine-tune, and deploy.

View on GitHub Weights on Hugging Face →
LUNA logoLUNA
Topology-agnostic transformer · 300× fewer FLOPsNeurIPS 2025
FEMBA logoFEMBA
Bidirectional Mamba · linear-time, 0.949 AUROCEMBC 2025
LuMamba logoLuMamba
LUNA + FEMBA + LeJEPA · 377× cheaper than LaBraMEUSIPCO 2026
PanLUNA logoPanLUNA
Multimodal EEG + ECG + PPG · 5.4M paramsAICAS 2026
TinyMyo logoTinyMyo
3.6M-param EMG model for microcontrollersarXiv 2025
CBCEReBrO
Compact encoder · alternating attentionarXiv 2025
Work with me

Open MSc & semester projects at ETH Zürich

I supervise students on tiny recursive models for time series and quantized TRMs on edge hardware — real hardware, weekly supervision, and a paper-shaped goal from day one. Recent student work landed at NeurIPS, EMBC, and IEEE journals.

Work with me → Open topics & how it works

Recent Posts

Thermo-TRM: Recursive Reasoning at Thermodynamic Prices
At the ETH Zurich Probabilistic Computing Hackathon I reframed the recursion in TRM/PTRM/GRAM as annealed sampling on an energy-based model, ran the whole loop on Extropic’s THRML simulator, and solved the hardest known Sudoku as a ground-state problem. The recursion buys roughly 7x fewer sampling sweeps, and when the network is confidently wrong the physics rescues it. It won 2nd place.
Thorir Mar Ingolfsson
Last updated on 25 Jun 2026 8 min read
Project
Thermo-TRM: Recursive Reasoning at Thermodynamic Prices
LuMamba: How the Pre-training Objective Shapes an EEG Foundation Model
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.
Thorir Mar Ingolfsson, Danaé Broustail
Last updated on 15 Jun 2026 10 min read
Project
LuMamba: How the Pre-training Objective Shapes an EEG Foundation Model
LUNA: An EEG Foundation Model That Doesn't Care How Many Electrodes You Have
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.
Thorir Mar Ingolfsson
Last updated on 7 May 2026 8 min read
Project
LUNA: An EEG Foundation Model That Doesn't Care How Many Electrodes You Have
See all posts

Recent Publications

Tip: Explore the full archive and filter by venue, topic, or year on the publications page.
Elisa Vasta, Thorir Mar Ingolfsson, Andrea Cossettini, Luca Benini, Tilman Beck, Emanuela Keller, Una Pale (2026). Evaluation of EEG Foundation Models for Event-Based Burst-Suppression Detection in ICU. arXiv preprint.

PDF Project DOI

Marija Zelic, Anna Tegon, Yawei Li, Thorir Mar Ingolfsson, Luca Benini (2026). PanLUNA: An Efficient and Robust Query-Unified Multimodal Model for Edge Biosignal Intelligence. In IEEE AICAS 2026.

PDF Project DOI

Danaé Broustail, Anna Tegon, Thorir Mar Ingolfsson, Yawei Li, Luca Benini (2026). LuMamba: Latent Unified Mamba for Electrode Topology-Invariant and Efficient EEG Modeling. In EUSIPCO 2026.

PDF Project DOI

Anna Tegon, Nicholas Lehmann, Yawei Li, Andrea Cossettini, Luca Benini, Thorir Mar Ingolfsson (2026). FEMBA on the Edge: Physiologically-Aware Pre-Training, Quantization, and Deployment of a Bidirectional Mamba EEG Foundation Model on an Ultra-Low Power Microcontroller. In IEEE TBME.

PDF Project DOI

Matteo Fasulo, Giusy Spacone, Thorir Mar Ingolfsson, Yawei Li, Luca Benini, Andrea Cossettini (2025). TinyMyo: A Tiny Foundation Model for Flexible EMG Signal Processing at the Edge. arXiv preprint.

PDF Project DOI

See all publications

Contact