Can a computer read
your motor
intentions?

Imagining a movement
fires up your motor cortex.

Your brain stores a kind of internal simulator for every movement you know. When you actually move, it runs the simulator and sends a go signal to your muscles. When you imagine moving, it runs the same simulator but holds the go signal back. The planning machinery still fires up. No movement happens, but the brain activity is real.

That is the entire basis of motor imagery BCI. You do not need to move. You just need to think about moving, consistently, on cue. In this dataset, nine volunteers imagined squeezing their left hand, their right hand, their feet, and their tongue, one cue at a time, for 288 trials per session.

The signal is a dip,
not a spike.

EEG electrodes on the scalp pick up tiny voltage shifts from thousands of neurons firing in sync below. At rest, the motor cortex ticks along with a steady rhythm, like a crowd clapping together. When you prepare or imagine a movement, that rhythm breaks. Neurons stop clapping in sync and start doing their own thing. On a frequency plot, you see a drop in power in the 8 to 30 Hz range. That drop is called event-related desynchronization, or ERD.

Here is the part that makes classification possible: the dip happens on the opposite side of the brain from the imagined limb. Imagine squeezing your right hand, and the signal drops under electrode C3 on the left hemisphere. Imagine your left hand, and it drops under C4 on the right. That left-right asymmetry is the signal a classifier reads.

The algorithm finds the
best angle to separate classes.

You have 22 electrodes and want to tell left hand from right hand. Most of the signal is noise. A few electrode combinations are diagnostic. Common Spatial Pattern filtering asks: what weighted mix of all 22 channels makes the signal as loud as possible for class A while making it as quiet as possible for class B?

Think of 22 microphones spread across a noisy room. CSP figures out how to blend them so one imagined class rings clearly and the other fades. After filtering, a simple linear classifier reads the power in those blended channels and outputs a prediction: left hand, right hand, feet, or tongue.

A smarter way to measure
brain state distance.

CSP + LDA finds a linear combination of electrodes that separates two classes. It works. But it makes assumptions: that the useful information lives in the variance of individual channel combinations, and that two covariance matrices that look similar numerically represent similar brain states.

Riemannian geometry takes a different view. Instead of treating each EEG epoch as numbers to be linearly combined, it treats each covariance matrix as a point in a curved geometric space. Distance between brain states is computed along the surface of that space, not in a straight line through it. The classifier assigns a trial to whichever class centroid it is closest to, measured in that curved distance. In practice, this tends to be more robust to noise and to the signal amplitude differences you see between subjects.

Browse the signals.
Watch the classifier work.

This tool lets you explore the full pipeline, from raw EEG to classifier output, in your browser. No Python environment. No downloads. Start with the signals, run CSP + LDA on any subject, then compare it head to head against Riemannian geometry on the same data. Phase 2 adds EEGNet and deep learning models to the comparison.