Read intent
Turn signals into decisions. The three BCI paradigms, real experiments, and reproducing published work.
An amplifier gives you a trace. A decoder gives you a decision. This phase builds the three classic brain-computer interfaces in order of difficulty, teaches you to run a clean experiment on another person, and has you reproduce a published result and explain why your number differs from theirs.
Three paradigms, three difficulty levels. SSVEP works in an afternoon and feels like magic. P300 takes a weekend and teaches you averaging and classification. Motor imagery takes months and teaches you why the field is not finished. Do them in that order.
Alongside the paradigms, two things that most engineering students never do: run a real human experiment, with timing, consent, and a protocol, and reproduce a published result number for number. The second one is where you learn that most published accuracies are a little too good, and why.
A warning about honesty
The most common flaw in student BCI work is data leakage: trials from the same session, or the same person, landing in both training and test sets. Accuracies inflate by ten to thirty points. The explainer on classification is mostly about not doing this. Read it before you report a number to anyone.
Projects
Choose one. Do a second if hooked.Project A: The SSVEP speller
Boxes flicker at different rates; the back of your head flickers with whichever one you look at. A decoder reads the frequency out of the EEG and types the letter. Works in an afternoon, and it is the BCI that actually works.
ProjectProject B: The P300 speller
A grid of letters flashes row by row and column by column. The one you are counting produces a bump 300 milliseconds later. Average, classify, and the computer types it. The 1988 design that started the field, built in a weekend.
ProjecthumblingProject C: Motor imagery
Imagine moving your left hand or your right. A classifier tries to tell which from rhythms over motor cortex. Months of work, subject-dependent, and the project that teaches you why BCI is not solved.
ProjectlabProject D: Your own ERPs
P300, the face-sensitive N170, and the error-related negativity, recorded from friends under a lab's protocol with PsychoPy and LSL. Learning to run a clean human experiment is a skill most engineers never acquire.
ProjectProject E: The sleep band
Two forehead electrodes, an automatic sleep stager trained on public data, and then pink-noise pulses timed to slow waves, which has published effects on memory. A freshman who builds a closed-loop sleep device has a story no one else has.
ProjectProject F: Reproduce a paper
Download the data behind a published result, reimplement the method, get a number, and explain why it differs from theirs. A graduate-level portfolio piece built from a laptop, and the fastest way to learn what published accuracies mean.
ProjectProject G: Neurofeedback, and how to test whether it does anything
Build an alpha trainer that shows your own rhythm in real time. Then read the skeptical literature, design a sham-controlled test, and find out whether your device does anything. The critical thinking is the product.
ProjectProject H: Play
Alpha tug-of-war for two, live EEG sonification, a brain-controlled lighting rig. Underrated for learning, and the best recruiting demos a NeurotechX club can have.
Explainers
Theory that attaches to what you just built. Read after doing.Event-related potentials and oscillations
Two ways the brain's electrical activity carries information: a stereotyped waveform locked to an event, and a rhythm whose power or phase changes. Averaging, time-frequency analysis, and the difference between evoked and induced.
ExplainerClassification done honestly
Features, LDA, CSP, and Riemannian classifiers, then the part that matters more than the algorithm, which is how you split the data. Leakage, chance levels, calibration drift, and information transfer rate.
ExplainerBayesian decoding and the Kalman filter
A decoder as an estimator that blends what it expected with what it measured, weighted by how much it trusts each. The population vector, the Kalman filter, and why intracortical cursor control works.
ExplainerDeep learning for EEG
EEGNet and its relatives, why they are small, what data augmentation does, what foundation models promise, and the honest benchmark result that deep networks often tie the Riemannian baseline on EEG-sized data.
ExplainerDesigning an experiment that can fail
Hypotheses, counterbalancing, jitter, triggers, sample size, pre-registration, consent, and the pilot. The parts of a human experiment that decide whether the result means anything, before any data exist.
ExplainerStatistics for small experiments
The binomial bound that says when a classifier is above chance, permutation tests that need no assumptions, cluster correction for waveforms, effect sizes, and the multiple-comparisons trap. Enough to not fool yourself.
Ready for Phase 4 when
Checks save on this deviceWhat an interviewer would ask
If you can answer these, you are done here- You report 92 percent accuracy on four-class motor imagery. What is the first question I should ask you?
- Explain a Kalman filter to a first-year student using a cursor as the example.
- Why do deep networks so often fail to beat Riemannian classifiers on EEG datasets?
- What is the information transfer rate of a speller, and why is accuracy alone misleading?