Skip to content

Non-Invasive Recording

Non-invasive BCIs acquire signals from the scalp, outside the skull, or via external devices — no craniotomy, simpler regulation, lower cost, but signal quality is markedly lower than invasive approaches. This is the only realistic option for consumer-grade BCI and large-scale neuroscience research.

1. Four Major Non-Invasive Technologies

Technology Measured quantity Temporal resolution Spatial resolution Portable Cost
EEG Scalp potential ms 2–3 cm Very high \(100–\)5000
MEG Brain magnetic field ms 5 mm Low (shielded room) $$$$$
fMRI BOLD (oxygenation) ~1 s 1–3 mm (whole brain) None $$$$$
fNIRS Blood-oxygen spectrum ~100 ms ~1 cm Medium $$

2. EEG (Electroencephalography)

Electrode configurations

  • International 10–20 system: 21 electrodes standard; extended to 10–10 (64/128 channels)
  • Dry vs. wet electrodes: Dry electrodes are portable (Muse, Emotiv); wet electrodes offer better SNR (research-grade Biosemi ActiveTwo)
  • Sampling rate: 256–2048 Hz

Bandwidth and components

Component Frequency / trigger Feature
δ/θ/α/β/γ Frequency band State monitoring, imagery intent
P300 Event-related (300 ms) Response to rare stimulus
SSVEP Steady-state visual (6–30 Hz) Visual-flicker synchronous response
μ rhythm 8–12 Hz sensorimotor Motor-imagery suppression
N400 400 ms Semantic anomaly

Typical BCI paradigms

  • P300 speller (Farwell-Donchin 1988): Character matrix flashes; EEG decodes the fixated character
  • SSVEP speller (ITR up to 300 bits/min)
  • Motor imagery (MI-BCI): Imagined left/right-hand movement → μ rhythm suppression → 2D cursor

Limits

Skull low-pass filtering + spatial blurring make it hard for EEG to decode: - Fine finger movements (insufficient spatial resolution) - Word-level speech (insufficient bandwidth) - Complex continuous actions

Meta's 2023 pivot from EEG to MEG was driven precisely by EEG's signal-precision ceiling.

3. MEG (Magnetoencephalography)

Principle

Neural currents produce tiny magnetic fields (~10 fT, 10⁸ times weaker than Earth's magnetic field); detected by SQUID (superconducting quantum interference device) sensors.

Advantages

  • Unaffected by skull attenuation (magnetic fields barely attenuated by skull)
  • Spatial resolution ~5 mm (5–10× better than EEG)
  • Millisecond temporal resolution

Disadvantages

  • Requires a magnetically shielded room (to eliminate Earth's field and environmental interference)
  • Requires liquid-helium-cooled SQUIDs (in recent years OPM — optically pumped magnetometers — have begun to replace them, requiring no liquid helium and enabling helmet-style devices)
  • A single unit starts at $2M

Representative research

Meta AI / Défossez et al. (2023) decoded heard speech from MEG: - Within a 3-second window, identified the heard word from ~1500 candidates - 41% top-10 accuracy - Purely non-invasive, no training required (using pretrained speech representations)

Significance: Demonstrates that non-invasive brain-to-speech is feasible, but the data type is auditory perception (user "listens"), not "active thinking."

4. fMRI

Principle

BOLD (Blood Oxygen Level Dependent) signal: neural activity → local blood-flow increase → oxygenated hemoglobin change → MR signal change.

Features

  • Poor temporal resolution (~1 s, with 4–6 s hemodynamic response delay)
  • Excellent spatial resolution (1–3 mm, whole brain)
  • Not portable (3T/7T scanner)

Representative BCI research

fMRI is the workhorse for decoding semantics and visual content:

  • Takagi & Nishimoto 2023: Stable Diffusion + fMRI reconstructs perceived images
  • MindEye / MindEye2 (2023/2024): CLIP + diffusion reconstructs natural images, reaching human-level recognition
  • Tang et al. 2023 Nat Neurosci: GPT-based semantic decoding of "the gist of heard stories"

fMRI suits thought / semantic decoding rather than real-time control — hemodynamic latency rules out real-time closed-loop BCI.

5. fNIRS

Functional near-infrared spectroscopy uses near-infrared light (700–900 nm) to penetrate the skull and measure blood oxygenation:

  • More portable than fMRI (helmet form factor)
  • Better spatial resolution than EEG
  • Temporal resolution ~100 ms (better than fMRI, worse than EEG)
  • Penetration depth only ~3 cm; can sample only superficial cortex

Kernel Flow is the consumer-grade fNIRS representative (released 2022). Meta (Facebook Reality Labs) invested in fNIRS BCI from 2019–2021, then pivoted to EMG in 2021.

6. EMG: The "Fifth Pillar" of Non-Invasive

Strictly speaking, EMG (electromyography) is not a brain signal, but it is very popular in consumer-grade "silent speech" / "gesture BCI" applications:

  • Meta Reality Labs Orion (2024 demo): Wristband EMG for intention-level gesture recognition
  • CTRL-Labs (acquired by Meta in 2019) EMG wristband
  • Silent speech: Detect throat/facial EMG to infer unspoken words (MIT AlterEgo 2018)

EMG advantage: far higher precision than EEG (muscle signals are ~100× stronger than brain signals); but it is not "reading the brain."

7. Information-Theoretic Ceiling of Non-Invasive BCI

Under the Wolpaw ITR (information transfer rate) framework, the limits of non-invasive BCI are:

  • Typical EEG SSVEP: 100–300 bits/min (close to typing speed)
  • Motor-imagery EEG: 20–30 bits/min (cursor control)
  • MEG speech: about 10–20 bits/min (word-level recognition)
  • Invasive spike (Willett 2023): ~62 WPM ≈ 1800 bits/min

Conclusion: On high-ITR tasks, non-invasive BCI lags invasive by 50–100×. This is why consumer and clinical tiers are entirely different product categories.

8. The AI Frontier of Non-Invasive BCI

Recent deep learning has significantly improved non-invasive BCI:

  • EEGNet (Lawhern 2018): General-purpose EEG CNN architecture
  • EEGPT (Pu 2024): Million-scale pretrained EEG Transformer
  • Meta MEGFormer 2023: MEG speech-decoding foundation model
  • DeWave (Duan 2024): EEG → discrete tokens → LLM for semantic reconstruction

Common pattern across these works: self-supervised pretraining + downstream fine-tuning — transferring the NLP / CV foundation-model paradigm to non-invasive BCI.

9. Logical Chain

  1. Non-invasive BCI = consumer-grade BCI; EEG is the only non-invasive route commercialized at scale.
  2. fMRI suits semantic decoding, MEG suits fast decoding, and fNIRS suits portable state monitoring.
  3. EMG is the alternative path chosen by Meta and others — skip "reading the brain" and go straight to gesture BCI.
  4. Non-invasive ITR is 50–100× lower than invasive, making the two tiers different products.
  5. EEG foundation models + LLM fusion is the AI breakthrough for non-invasive BCI to close the gap with invasive.

References

  • Wolpaw & Wolpaw (2012). Brain-Computer Interfaces: Principles and Practice. Oxford. — Authoritative textbook on non-invasive BCI
  • Défossez et al. (2023). Decoding speech perception from non-invasive brain recordings. Nat Machine Intelligence. https://www.nature.com/articles/s42256-023-00714-5
  • Takagi & Nishimoto (2023). High-resolution image reconstruction with latent diffusion models from human brain activity. CVPR.
  • Tang et al. (2023). Semantic reconstruction of continuous language from non-invasive brain recordings. Nat Neuroscience. https://www.nature.com/articles/s41593-023-01304-9
  • Lawhern et al. (2018). EEGNet: a compact convolutional neural network for EEG-based brain-computer interfaces. J Neural Eng.

评论 #