Record the brain
EEG measures electrical signals through electrodes
Record movement
A small sensor attached to the electrode measures how it moves, which can help explain noise in the brain recording
Guess the label
For each two-second section: was the electrode deliberately disturbed, or left undisturbed?
Why care? I wanted to explore what makes brain devices useful in everyday life. As a first step, I tested whether movement measurements help a small model recognize deliberate electrode disturbance in a lab, then hid those measurements to see how its answers changed.
Both labels refer to a recording that is still running. “Left undisturbed” means the electrode was not deliberately moved during that section. This experiment studies electrode disturbance, rather than walking or head movements.
Does movement help?
Compare brain data alone with brain data plus measurements of how the electrode moved.
What if it disappears?
Hide the movement reading at test time and see how the score changes.
Can we trust a confident answer?
Keep only the answers the model feels sure about and see what gets left out.
What did I find?
The original test showed no clear benefit from adding movement. In the follow-up using the difference between two movement readings, hiding movement dropped the score from 68.1% to 62.7%. Practising with missing readings brought it back to 66.7%, about the same as EEG alone. That makes the simple EEG-only model a useful comparison for a more complicated system.
The follow-up was chosen after seeing the first results and needs a fresh test. These recordings also do not establish performance on new people or during everyday activity.
How does this connect to larger brain models?
Models that learn from lots of brain recordings still need useful measurements and careful testing. This project starts with a small, concrete question: does an extra measurement help when it is present, and can we cope when it is missing?
I tested small models here. These results do not establish how a foundation model would perform. BrainLM is one example of research on larger brain models.
Does the movement reading help?
Here, “movement” is the electrode’s movement, measured by a small sensor attached to it. Researchers deliberately disturbed one electrode while another was left alone.
What makes these four choices different?
- Brain + movement: learns how much weight to give each measured feature.
- With missing-data practice: learns the same way, but also sees training examples with movement hidden.
- Using decision rules: learns a series of if-then checks, where each result decides which measurement to check next.
- Brain recording only: learns without movement.
3. What can it see during the test?
The brain recording stays available. This switch hides movement from an already trained model.
Score on recordings kept out of training
One input, or both together?
Each bar is a separately trained model. The combined models use both kinds of measurements.
Brain measurements+Movement measurements→Train a combined model
Detection score, higher is better
What does the score mean?
The model labels each two-second recording as sensor left undisturbed or sensor deliberately disturbed. The score averages how often it gets each kind right, giving both kinds equal importance, then averages across recordings.
Randomly choosing between the two labels would score about 50%. That is a comparison point, not the model’s current score.
More sure. More right. But fewer answers?
The model attaches a probability to each answer. Here we keep answers only when that number reaches your chosen cutoff, then check them against the recorded labels.
50% · keep all70%90% · more selective
Each block is about 1% of the examples.
Dark green: answered. Pale green: skipped.
of all examples get an answer.
Why does 50% keep every answer?
There are two possible labels. Their probabilities add up to 100%, and the model chooses the larger one. So its chosen answer always has at least 50%. A 70% cutoff means it skips answers below 70%; it does not promise to get 70% right.
The practical question: Is it answering often enough, in the situations we care about? A nearly perfect score can hide a model that skips most recordings.
This page counts correct answers among the examples kept. The sensor page uses a different score that gives undisturbed and deliberately disturbed periods equal weight. These cutoffs were explored after testing, and have not been established as safe settings for a device.
How I checked the idea
1. Prepare recordings
Split 23 recordings into 5,642 two-second examples, each labelled sensor left undisturbed or sensor deliberately disturbed.
2. Learn patterns
Describe signal size, changes and rhythms. Train models using some recordings.
3. Test on the rest
Keep entire recordings out of training. Compare predictions with their labels.
This is important as remembering an example is easier than answering on a recording the model has not seen. Recordings from the same person may still appear on both sides, so this does not prove it works for new people.
Check the simpler option
Does adding movement beat using the brain recording alone?
Take information away
Does practice with missing data help when movement disappears?
Look beyond the headline
Which examples get skipped when we only keep confident answers?
Why are there two movement measurements?
I first used the movement reading at one location. Adding it gave no clear improvement over brain data alone.
After seeing that result, I tried subtracting readings from two locations. This describes their difference in movement and assumes the sensors’ axes can be compared. It showed a small improvement, but needs a fresh test because I chose it after the first analysis.
How did I check whether a small lead could be luck?
I repeatedly drew groups from the saved recording results and recalculated the difference between models. If those comparisons include both a gain and a loss, the evidence does not point to a clear winner. Repeated recordings from the same people remain a limitation.
What would I test next?
New people, natural everyday movement, and a fresh test of the two-location idea. The source experiment deliberately disturbed sensors, so it cannot establish performance during ordinary daily activity.