TTechnology

What the sensors can see, and what a model can do with it.

Two layers. The signal layer is physics and physiology: what reaches an electrode and in what state. The model layer is what can be recovered from it. Conflating the two is how neurotechnology gets oversold.

01Signal layer

EEGCortical activity, measured through skull and skinElectrodes on the scalp pick up summed electrical activity from large populations of neurons. By the time it reaches the surface it is small — commonly tens of microvolts — spatially blurred by the tissue in between, and sitting in a world of larger electrical noise.

EMGMuscle activity, much larger and much closerJaw, facial and neck muscles generate their own electrical activity, typically far larger than EEG and concentrated in higher frequencies. Near the ear it is unavoidable. Whether it is only interference, or also a source of intent, is an open question Cerebral treats as a question rather than a settled matter.

Around-earDiscreet placement, heavily reduced coveragePutting electrodes in and around the ear buys comfort, discretion and the possibility of all-day wear. It costs spatial coverage: a small number of closely spaced contacts, positioned far from much of the cortex, with a different noise profile from a scalp montage.

The mixtureWhat actually reaches the amplifierA real recording is a superposition: cortical activity, muscle activity, eye movement, cardiac artefact, electrode and cable movement, and mains interference. Separating these is a large part of the problem, not a preprocessing footnote.

Amplitude ranges above are general background from the EEG literature, not Cerebral measurements. Exact figures depend on electrode type, reference, montage and task.

02What a decoder is, precisely

A decoder maps a short window of multichannel voltage onto a label. That is the whole operation. It is not reading language, and it is not reading thought.

Different BCI tasks are routinely spoken about as if they were one capability. They are not. Silent speech, attempted articulation, motor imagery, evoked-potential selection and cognitive-state estimation differ in what the user does, what the signal contains, how much training data is needed, and how well any of it transfers to a new person. Cerebral names the task wherever it makes a claim.

03Models

Every significant computational model gets a model card with the same fields, including the uncomfortable ones.

Model cards and experiments

04Interactive demos

PlannedAn electrode-configuration explorer, an annotated signal viewer, a performance-versus-sensor-count chart and a model comparison. None of them can exist honestly before there is data to put in them.