Low-channel-count transfer learning
Can models pre-trained on large public EEG datasets be adapted to 8 or 16 consumer-grade channels? A good topic for a thesis or a preprint.
Possible Outcome
Code, model weights and a preprint (e.g. arXiv).
You own what you build. Publish it under your own GitHub account with your name and chosen open-source licence.
Why It Matters for Open Neurotech
Foundation models for EEG are emerging rapidly, but most assume 64+ wet lab channels. Bridging foundation models down to 8–16 channels with minimal per-user calibration unlocks practical everyday BCIs.
Starting Points & Technical Notes
These are open-ended suggestions, not specifications or assigned tasks. Change the scope, combine ideas, or take a completely different approach:
- Sub-select standard 10–20 subsets (8ch / 16ch) from PhysioNet, MOABB, or TUH EEG datasets for pre-training.
- Quantify how many minutes of target-user calibration data are needed for fine-tuning vs. training from scratch.
- Release model checkpoints and evaluation scripts under an open license.
Independent Open-Source Work
PiEEG Ltd does not take on volunteers, interns or unpaid workers. Any project you start is your own independent activity under your own repository and responsibility. Read the full ground rules →
Safety First
PiEEG devices are not medical devices. Use battery power only when electrodes are on a person, and never connect to mains power during recording. Liability & safety →
More Signal Processing & ML Ideas
Artifact detection benchmark
Compare classic and ML methods for removing blinks, jaw clenches and motion artifacts from low-channel EEG, using public datasets or your own recordings.
Real-time motor-imagery decoder
Build an open, real-time left/right motor-imagery classifier that runs on a Raspberry Pi or laptop and streams from an open EEG board.
SSVEP speller in the browser
A web-based SSVEP speller with flicker targets and CCA detection that works with an open EEG data stream.