← All projectsPage 2 · Continued
Neural Decoding · In progress · 2026
Decoding imagined speech from EEG
Can phrases said silently in the head be told apart from scalp EEG? A four-person capstone built to measure a modest effect honestly rather than report an inflated one.
Evaluation framework built · first real experiments next · no results
The question
Imagined speech is a word said silently in the head, with no movement at all. Decoding it from electrodes on the scalp is hard, and the field is full of near-perfect accuracies that come from leaks between training and test data rather than from the brain. We set the bar where it belongs: 55–70% on two to five classes would be a real result, and anything above 90% on three or more classes is treated as a bug until proven otherwise.
The data
Public datasets only.
| Dataset | Task | Role |
|---|---|---|
| BCI Competition 2020, Track 3 | Five imagined phrases (Hello, Help me, Stop, Thank you, Yes) · 15 subjects · 64 channels | Main target, with its official train, validation and test splits |
| Nieto et al. 2022 | Four imagined directions · 10 subjects, 3 sessions each · 128 channels | Second dataset, and the only one recorded over several sessions |
Rules before results
- Split by trial. Windows cut from one trial never land on both sides of a split, and every split is checked for it automatically.
- Fit inside the fold. Anything that learns from data, from normalisation to feature selection, is fitted on training data only.
- Three protocols, reported separately. Within one subject, across sessions, and on a subject the model has never seen. The drop between them is a finding.
- Honest chance. Not 1/k: the binomial threshold for the actual test size, plus permutation tests. On Track 3, with 50 test trials per subject, chance is 30%, not 20%.
- Every real run is logged, with its commit, protocol, seed, test size and chance threshold beside the accuracy.
Where it stands
The evaluation framework is built: loaders for both datasets behind one data format, leakage-checked splitters for all three protocols, the chance-level and permutation tests, a majority-class baseline and a Riemannian-geometry baseline, with tests and CI on every pull request. No model has been run on real data yet, so there are no numbers to show.
Team
Four-person capstone. My part: TK