Theses›Datacentre & AI compute›ML closes the signal-to-noise gap on non-invasive BCI

ML closes the signal-to-noise gap on non-invasive BCI

The standing bottleneck on non-invasive brain-computer interfaces is signal-to-noise: scalp EEG, MEG, fNIRS and functional ultrasound are noisy, low-spatial-resolution proxies for neural activity, which is why useful BCI has needed surgery.

  • Not dated

The call, a draftThe noise is winning. Machine learning has improved decoding within one person but not across people, and the best results still need MEG in a shielded room. The window is early.

D1D2D3R1R2R3C1C2foragainst D1D2D3R1R2R3C1C2foragainst
Fig. 1The balance. Drivers and catalysts against restraints, each block sized by its weight; catalysts hang dashed because they have not happened yet, sized by their odds. It tips toward the restraints.
  1. D1Within-subject decoding is improving: EEG image decoding roughly doubled in a year to about 55% top-1, and Brain2Qwerty v2 runs in real time on MEG at 39% word error with 10 hours per subject.
  2. D2Wearable optically pumped magnetometers are taking MEG out of the shielded room and moving toward clinical use.
  3. D3Functional ultrasound reaches 5-8cm deep at about 100 micron resolution from a small transducer, and a 2025 chemical approach let about 94% of sound through human skull.
  4. R1Decoding does not transfer across people: the same EEG task falls to about 15% across subjects, and EEG is already near its information limit of about 175 bits a minute.
  5. R2The skull blocks depth: EEG, fNIRS and OPM-MEG only see the cortical surface, and deep ultrasound in adults still needs a skull window or an injected contrast agent.
  6. R3EEG foundation models often fail to beat compact task-specific networks in low-data settings, so scale does not reliably help.
  7. C1A peer-reviewed non-invasive interface reports bandwidth within 10x of an implant on a comparable task, credited to a machine-learning decoder.31 Dec 2028
  8. C2A wearable OPM-MEG system gets a clinical clearance or a multi-site commercial deployment.31 Dec 2028
What would prove itdated tests, open and resolved
20272028202920302031today123456789predictionswatch 20272028202920302031today123456789predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1The wearable cap claiming 30 words a minute, and whether it publishes accuracy on users it was not trained on.Dec 2026 · open
  2. 2Merge Labs announces the start of a human study or a named clinical partner between 6 Nov 2026 and 31 Jan 2027.Jan 2027 · open
  3. 3By end-2028 a peer-reviewed non-invasive BCI result (EEG/MEG/OPM/fNIRS) reports communication or control bandwidth within an order of magnitude of an invasive baseline on a comparable task, with the gain attributed to an ML decoder or learned denoiser.Dec 2028 · open
  4. 4Wearable OPM-MEG reaches a named clinical or multi-site commercial deployment milestone by 2028.Dec 2028 · open
  5. 5A non-invasive result within 10x of an implant, credited to a learned decoder.Dec 2028 · open
  6. 6A wearable OPM-MEG clinical clearance or multi-site deployment.Dec 2028 · open
  7. 7By 2030 the value-capture point in non-invasive BCI is the decoding/signal layer or the sensor front-end, not the headset brand: at least one non-invasive-neurotech company whose moat is the ML stack or the sensor reaches a >$500M outcome.Dec 2030 · open
  8. 8Consumer (non-clinical) non-invasive BCI does NOT reach mass-market scale (>1M active units) by 2030; SNR and reliability stay below the bar for general use.Dec 2030 · open
  9. 9Whether any consumer non-invasive interface passes 1M active units.Dec 2030 · open
Expressionslisted, private, technologies
Private · 2 private companies, 2 of them in the US.

Piramidal · Constellation

Technologies