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
- Non-invasive BCI = consumer-grade BCI; EEG is the only non-invasive route commercialized at scale.
- fMRI suits semantic decoding, MEG suits fast decoding, and fNIRS suits portable state monitoring.
- EMG is the alternative path chosen by Meta and others — skip "reading the brain" and go straight to gesture BCI.
- Non-invasive ITR is 50–100× lower than invasive, making the two tiers different products.
- 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.