Neural Intent Lab

The first result did not generalize.Subject 1 scored 66.7%. Across 20 participant-specific models, the mean fell to 56.3%.

What is actually happening?

We are testing whether recorded EEG can identify one of two imagined movements.

A person was shown either “left hand” or “right hand,” imagined that movement, and an EEG cap recorded the electrical activity across their scalp.

1Record

We use public EEG recordings from 20 people. Each recording already has the correct left/right answer attached.

2Calibrate

We train a separate model for each person using 30 labeled examples. Their model is not shared with other people.

3Hide the answer

The model receives 15 new examples from a different recording session without seeing their labels.

4Check

We compare its left/right guesses with the hidden answers. It got 169 of 300 correct: 56.3%.

What this provesThe software can train and predict from unseen EEG recordings.

What this does not proveIt cannot read words, thoughts, or operate reliably from a live headset.

Mean participant accuracy

56.3%

Only 6.3 percentage points above a two-class coin flip.

The larger test changed the conclusion.

Each person received a separate model trained on runs 4 and 8. Run 12 stayed untouched until evaluation. The pooled result was 169 correct predictions from 300 trials.

Not dependable enough for control.

The model made 105 wrong predictions at 80% confidence or higher. Confidence cannot rescue this pipeline.

Participant distribution

Performance varies dramatically by person

15 test trials each

S001

66.7%

S002

66.7%

S003

53.3%

S004

46.7%

S005

40.0%

S006

66.7%

S007

100.0%

S008

33.3%

S009

60.0%

S010

46.7%

S011

60.0%

S012

73.3%

S013

53.3%

S014

60.0%

S015

60.0%

S016

53.3%

S017

40.0%

S018

40.0%

S019

53.3%

S020

53.3%

Replay-first live pipeline

Watch held-out EEG become an intent token

The local app streams run 12 one EEG window at a time through the personalized model, emits a confidence-gated CONFIRM or CANCEL event, and reveals the hidden label afterward. Its second tab still accepts compatible EEGMMIDB EDF files. Everything stays on this computer.

Windows launcherresearch\eeg\run_predictor.bat

Recorded source now. The same EEGWindow boundary accepts a future headset adapter.

Inspect Subject 1

Held-out result viewer

Stored outputs · not live inference

This web viewer reveals the recorded predictions from Subject 1's untouched run 12. Use the local EDF predictor above for actual file inference.

The true label remains hidden until reveal.

Next scientific gate

Improve cross-session reliability before adding more commands.

Next we should compare stronger spatial models, calibrate confidence, add idle periods, and test continuous false activations. Four commands would currently compound an unstable two-command decoder.