ICA Component Gallery
A drill on at least 40 real ICA components — classify each from its topography, time course, spectrum and ERP image, then read the evidence.
3 claims on this page are unverified. TODO(confirm) marks a specific statement the author has not yet checked against a
primary source. Everything else on this page has been reviewed. Treat a marked claim as
provisional and go to the cited source rather than quoting the sentence.
Modes: drill — the widget below runs in drill. Use Share state to put the exact view in the URL.
What it does
Each item is one real independent component shown through four views at once: its scalp topography (drawn from the component’s channel weights and the shared spherical-spline interpolation matrix), a 10-second stretch of its time course, its power spectrum, and an ERP-image thumbnail of its single trials. You classify it as brain, eye, muscle, heart, line noise, channel noise or other; the reveal gives the reference label, the evidence behind it, and the ICLabel probabilities where the classifier was available.
Everything is precomputed. The decomposition, the spectra, the ERP images and the labels are produced in Python and shipped; the browser draws them and scores the drill.
Controls
| Control | What it sets |
|---|---|
| Class buttons | Brain · eye · muscle · heart · line noise · channel noise · other |
| Hint | One view at a time, so you can practise using the evidence in a chosen order |
| Reveal | Reference label, rationale, the evidence list, and ICLabel probabilities if present |
| View toggles | Show or hide topography, time course, spectrum and ERP image individually |
| Next / previous | Move through the session; the score is reported to the exercise when the session ends |
What to look for
drill
- A class is identified from the convergence of views, never from one. Force yourself to name the evidence before pressing a button.
- Topography: a blink is a broad frontal-polar field with the same polarity on both sides; a horizontal saccade is lateral frontal with opposite polarity across the midline; muscle is focal at the rim with a sharp, non-dipolar edge; channel noise is one hot electrode with nothing around it; a brain component is smooth and dipolar.
- Spectrum: brain components usually show a 1/f-like background with a peak (alpha near 10 Hz, mu with its arch harmonic); muscle power rises with frequency; line noise is a single spike at the mains frequency and its harmonics; a blink component’s power sits below a few hertz.
- ERP image: a brain component carrying the effect shows a deflection aligned to the event across trials; a blink component shows bright bands at whatever latency the participant blinked, often after the response.
- The ambiguous components are the lesson. When two views disagree, decide which one the physics supports — and notice that the cost of being wrong is not symmetric between classes.
Used in
- L2.6 EOG regression and ICA (
drill)
Data provenance
ds-erpcore — ERP CORE (Kappenman et al., 2021), CC BY 4.0, open access, per-subject downloadable. Biosemi ActiveTwo, 30 EEG + 3 EOG electrodes, 1024 Hz, CMS reference, 60 Hz mains, no software filters. At least 40 components from one subject’s P3 recording, fitted on a 1 Hz-filtered copy with the rank the recording actually has and a recorded seed (L2.4, L2.6); the shipped products are derived (filtered, decomposed, epoched, resampled) and attributed.
Labels carry label_source: algorithmic (§4.5) until the author reviews them, and each item ships the rationale and evidence that produced its label. Where mne-icalabel was available, per-class probabilities ship alongside; where it was not, the labels come from documented feature heuristics and gallery.json and the widget README say so.
TODO(confirm): the subject id, the exact number of components, the ICA method and seed, and whether ICLabel probabilities shipped are read from gallery.json and data/manifest.json once data/scripts/make_ica_gallery.py has run; the frame’s provenance line is the authority. TODO(confirm): the author mirrors the ERP CORE entry into data/registry.yaml (§10.11 item 8), and promotes the component labels to label_source: expert after review — until then the drill grades against a documented procedure, not a person.
Open the code
site/src/components/widgets/w-ica-component-gallery/ — Widget.svelte, index.ts, README.md. Repository link: TODO(confirm) (GitHub org/repo, §13 item 3).