Every trial is a short stretch of EEG after a picture appears: 10 electrodes, −200 to 800 ms. The data are synthetic, so the truth is known. Both conditions share the usual components (P1, N170, P3), but for faces the N170 — a negative deflection around 170 ms — is stronger at the posterior-lateral electrodes P7, P8, PO7 and PO8, and there is a smaller, later positivity around 400 ms at Cz and Pz. Everything else is noise, and a single trial is mostly noise.
A logistic regression is trained on 280 trials to predict the condition from all 510 values (electrode × time), then tested on 120 trials it has never seen. The heatmap is the attribution map: how much each electrode at each moment moved the model's decisions, averaged over the held-out trials. For a linear model the SHAP value of a feature is exactly weight × (value − average value), so nothing here is approximated. This page is a linear miniature of the idea, not the published pipeline.
Things to try: turn the effect size to zero — accuracy falls to chance and the map dissolves into speckle. Raise the noise and watch both degrade together. Tick "show planted difference" to check the map against the ground truth, which real data never give you. And a caveat that carries over to real data: the map shows what this model used, which is evidence about the brain only as far as the model is good.