Origins of Neural Signals
Different BCI electrodes observe signals at entirely different physical scales. This section explains the physical origins, spatial scales, and decoding upper bounds of the four signal types: spike, LFP, ECoG, and EEG. These determine what a BCI system can and cannot do.
1. Basics of Neuronal Electrical Activity
Neurons generate electrical signals by opening and closing ion channels. Two key events:
- Action potential (spike): An all-or-none pulse generated at the soma; waveform lasts 1–2 ms with ~100 mV intracellular amplitude. This is the nervous system's digital signal.
- Postsynaptic potential (PSP): A continuous potential change at the dendrite upon input; small amplitude (mV-range), long duration (10–100 ms). The superposition of many PSPs generates the local field potential.
External electrodes observe these events as potential changes in the extracellular space. As electrode distance grows, high-frequency components (spikes) attenuate significantly from current spreading and tissue filtering, while low-frequency components (LFP, ECoG, EEG) remain relatively intact.
2. Physical Origins of the Four Signals
| Signal | Physical origin | Recording distance | Bandwidth | Amplitude |
|---|---|---|---|---|
| Spike | Single-neuron action potential | < 100 μm | 300 Hz–6 kHz | 50–500 μV |
| LFP | Local neuron-population PSP and synchronous spikes | 0.1–1 mm | 1–300 Hz | 100 μV–2 mV |
| ECoG | Cortical-column population PSP | 1–5 mm | DC–300 Hz | 50–500 μV |
| EEG | Cortical synchronous PSP + skull attenuation | 1–3 cm | 0.5–100 Hz | 10–100 μV |
Core physical principles:
- More distant electrodes see larger, slower, and more synchronous signals. Because spatial distance attenuates high frequencies more than low, and more distant electrodes "see" more neurons — whose synchronous activity gets amplified by summation.
- Spatial and temporal resolution are roughly inversely related. Spike is single-cell + ms; EEG is cm-level but still ms-level (critical: EEG's temporal resolution matches spike's).
- Skull and CSF impose significant low-pass filtering and spatial smoothing on EEG. This is the root reason EEG decodable information is far less than ECoG.
3. Spike Signals
Physical nature
A spike is an action potential generated at the soma of a single neuron. In extracellular recording, neurons within the closest distance (< 50 μm) contribute most of the distinguishable spike waveforms.
Spike Sorting
Raw electrode signals are superpositions of multiple neurons' activity. Spike sorting separates different neurons' spikes:
- Detection: Threshold detection (events exceeding 3–5× noise SD)
- Feature extraction: PCA or wavelet transform to obtain waveform features
- Clustering: Algorithms like Gaussian mixture, Mountainsort, and Kilosort group similar waveforms into a single "unit"
Kilosort is the current mainstream tool. Neuropixels high-density probes have brought spike sorting into a new phase — separating hundreds of units simultaneously.
Decoding upper bound
Spike-level signals carry the highest information content:
- Fine finger movement (10+ DoF)
- Word-level speech (62 WPM, Willett 2023)
- 90 chars/min handwriting (Willett 2021)
- Complex robotic-arm control (7 DoF, Collinger 2013)
But invasiveness, tissue response, and long-term stability are the price.
4. LFP Signals
Physical nature
LFP (Local Field Potential) is the superposition of low-frequency electrical activity of neuron populations within the electrode's local range (0.1–1 mm). Main sources:
- Spatial summation of postsynaptic potentials (PSPs)
- Slow membrane currents
- Low-frequency components of partially synchronous spikes
Frequency bands
LFP is divided into frequency bands reflecting different processes:
| Band | Hz | Associated process |
|---|---|---|
| δ (delta) | 1–4 | Deep sleep, brain synchrony |
| θ (theta) | 4–8 | Hippocampal memory, spatial navigation |
| α (alpha) | 8–12 | Rest, closed-eye visual |
| β (beta) | 13–30 | Motor control, preparation |
| γ (gamma) | 30–80 | Attention, sensory binding |
| high-γ | 80–200 | Motor information close to spike level |
High-γ is the golden band for BCI: its information content approaches spike level, but its stability far exceeds spike (not affected by single-neuron loss from tissue response). Willett 2023's speech BCI mainly uses high-γ power features.
Decoding upper bound
LFP + high-γ achieves performance close to spike decoding, with far better long-term stability — this is the actual working bandwidth of most clinical BCIs outside of Neuralink.
5. ECoG Signals
Physical nature
ECoG (Electrocorticography) places electrodes subdurally or epidurally on the cortical surface. It observes cortical-column-level population activity (~10⁴ neurons).
- Spatial resolution: 1–5 mm (determined by electrode spacing)
- Temporal resolution: Same as LFP (DC–300 Hz)
- High-γ signal quality is close to LFP
Clinical use
ECoG has been used in clinical epilepsy monitoring for decades — the cortical recording method most thoroughly approved by the FDA. This makes ECoG BCI easier to clear: typical examples are Synchron Stentrode (strictly sEEG, but similar principle) and Precision Layer 7 (high-density thin-film ECoG).
Decoding upper bound
ECoG suffices for: - Speech BCI (Moses 2021, Metzger 2023 avatar are both ECoG) - Low-dimensional kinematic decoding - Emotional/affective state monitoring
But fine decoding down to single fingers or word-level speech boundaries still requires spike-level signals.
6. EEG Signals
Physical nature
EEG (Electroencephalography) records potentials from the scalp. Key challenge: skull resistivity is much higher than brain tissue, so the skull acts as a significant spatial low-pass filter on EEG. This degrades EEG spatial resolution to cm-level.
Common EEG components
| Component | Elicitation | BCI application |
|---|---|---|
| P300 | Rare stimulus (odd-ball) | P300 speller |
| SSVEP | Steady-state visual flicker (6–30 Hz) | SSVEP speller |
| Sensorimotor μ rhythm | Imagined hand/foot movement | 2D cursor control |
| ERP | Event-related potential | Cognitive-state detection |
| N400 | Semantic anomaly | Semantic-compatibility detection |
Decoding upper bound
EEG suits decoding of discrete choices and coarse-class motor imagery. Fine-grained speech, handwriting, and word-level semantic decoding exceed EEG's capabilities — which is why non-invasive works like Meta Défossez 2023 turned to MEG instead of EEG.
7. MEG and fMRI (Additional Comparison)
MEG
Magnetoencephalography records the magnetic fields generated by neural currents. It bypasses the skull's resistive suppression (magnetic fields are barely affected by skull), with spatial resolution ~5 mm, much better than EEG. But it requires a magnetically shielded room and SQUID sensors — expensive and immobile. Meta 2023 achieved 41% speech recognition with MEG.
fMRI
Functional MRI measures the BOLD (blood-oxygen-level-dependent) signal. Poor temporal resolution (~1 s), but spatial resolution of 1–3 mm with whole-brain coverage. Semantic and visual-content reconstruction (MindEye, Tang 2023) uses fMRI because of whole-brain coverage. But fMRI requires a scanner — non-portable.
8. Connection to Embodied-Intelligence Research
The physical scale of neural signals determines the working interface of different research communities:
- Computational neuroscience / world-model research: Spike + LFP provide signals closest to "computational primitives," supporting the Shenoy/Churchland dynamical-systems view.
- BCI engineering: ECoG is the most feasible for clinical translation and is the choice for most speech BCIs.
- AI-mind research (fMRI decoding): fMRI covers the whole brain and suits semantic-level research — the substrate for MindEye, Tang, and similar work.
- Consumer BCI: EEG is the only realistic choice, limiting intent complexity in consumer BCI.
9. Logical Chain
- Neural signals span four scales (spike / LFP / ECoG / EEG), each differing by 10–100× in spatial resolution and SNR.
- Signal scale bounds the granularity of intent decoding: spike/LFP for fine-grained decoding, EEG only for coarse classes.
- High-γ is the golden BCI band — information close to spike, stability close to LFP.
- ECoG is the best signal for clinical translation: fine enough, stable enough, surgical risk acceptable.
- MEG and fMRI each have their niches: MEG suits high-temporal-resolution semantic decoding, fMRI suits visual-content reconstruction.
- Different research communities use different signals to address different AI-science questions — the structural reason multiple communities advance BCI × AI in parallel.
References
- Buzsáki et al. (2012). The origin of extracellular fields and currents. Nat Rev Neurosci. — Authoritative review on neural-signal origins
- Crone et al. (1998). Functional mapping of human sensorimotor cortex with electrocorticographic spectral analysis. Brain. — Discovery of high-γ
- Pachitariu et al. (2016). Kilosort: realtime spike-sorting for extracellular electrophysiology. bioRxiv.
- Nunez & Srinivasan (2006). Electric Fields of the Brain: The Neurophysics of EEG. Oxford. — EEG physics
- Jun et al. (2017). Fully integrated silicon probes for high-density recording of neural activity. Nature. — Neuropixels