Skip to content

BCI Overview and Classification

1. What Is a Brain-Computer Interface

A Brain-Computer Interface (BCI) is a class of systems that establish an information pathway directly between the brain and external devices, bypassing peripheral nerves and muscles. Its core definition contains three elements:

  1. Signal source: Records from or stimulates the central nervous system (brain, spinal cord) directly — not peripheral physiological signals like EMG or eye movement.
  2. Information pathway: Transmits information bidirectionally between brain and machine — decoding (reading out) or encoding (writing in) neural activity.
  3. Purposefulness: Serves specific functions such as communication, control, sensory restoration, or neuromodulation — not pure recording.

A system belongs to the BCI category only if all three conditions hold. This definition excludes purely observational brain imaging (e.g., research fMRI scans) and peripheral-only devices (e.g., watch heart-rate + AI emotion inference).


2. Three-Dimensional Taxonomy

BCI systems can be classified along three orthogonal dimensions; together these dimensions capture all the key properties of a concrete BCI system.

Dimension 1: Invasiveness

This is the most direct classification dimension, determining spatial resolution, SNR, and surgical risk of the signal.

Category Electrode location Typical systems Spatial resolution Surgical risk
Fully invasive (Intracortical) Within cortex Utah array, Neuropixels, Neuralink N1 Single-neuron (~0.1 mm) High
Cortical surface (ECoG / sEEG) Subdural / cortical surface Clinical sEEG, Precision Layer 7 ~1 mm Medium
Minimally invasive Intravascular / epidural Synchron Stentrode ~few mm Low
Non-invasive Scalp EEG, MEG, fMRI, fNIRS cm (EEG) None

Signal quality increases monotonically along this sequence, but so does surgical risk. How to trade off "signal quality" against "surgical invasiveness" is the foundational problem of BCI engineering.

Dimension 2: Signal Direction

Direction Definition Typical applications
Read-out Decode intent or perception from neural activity Cursor control, speech BCI, handwriting BCI
Write-in Stimulate signals into the nervous system Intracortical microstimulation (ICMS), visual prosthesis, DBS
Bidirectional Simultaneous read and write Flesher 2021 arm with tactile feedback, Ganzer 2020 Cell

Traditional BCI focused on read-out; after 2016, "read-write closed-loop" represented by bidirectional BCI has become the new frontier (see Chapter 09).

Dimension 3: Application / Usage Scenario

Category User population Representative systems Regulatory pathway
Clinical BCI Paralysis, ALS, blindness, and other disabilities BrainGate, Neuralink, Synchron, Neuracle FDA IDE / NMPA Class III medical device
Neuromodulation Parkinson's, epilepsy, depression Medtronic DBS, NeuroPace RNS Multiple products on market
Research BCI Animal experiments, human experiments Neuropixels, university labs IRB / IACUC
Consumer BCI Healthy population Muse, Emotiv, OpenBCI FCC / CE
Augmentation / AR BCI Future vision No real commercial products yet No regulatory framework

This dimension best reflects the 2024-2026 changes: clinical BCI is moving from "single trial" into "multiple companies, multiple countries, parallel commercialization."


3. BCI Signal Hierarchy

Different acquisition methods observe signals at vastly different scales, which determines the upper bound of what can be decoded.

Signal Temporal resolution Spatial resolution Sampled object Typical channel count
Spike (single-neuron action potential) 1 ms Single cell 10² neurons 100–10k
LFP (local field potential) 1 ms 100 μm–1 mm 10³ neuron populations 100–1k
ECoG 1 ms 1 mm 10⁴ neurons 64–256
EEG 1 ms 1 cm 10⁶ neurons 8–256
MEG 1 ms 5 mm 10⁶ neurons ~300
fMRI ~1 s (BOLD delay) 1–3 mm 10⁶ neurons ~10⁵ voxels
fNIRS ~1 s 1–3 cm 10⁶ neurons ~50

Key pattern: the higher the spatial resolution, the finer the decodable intent granularity. Spike-level signals can decode word-level speech, complex handwriting, and fine-grained finger movement; EEG-level signals typically only support discrete choices (e.g., P300 speller) or large-amplitude motor imagery.


4. Active vs Passive BCI

Following Zander & Kothe 2011's taxonomy, BCIs admit an orthogonal division by the degree of user-intent engagement:

  • Active BCI: The user actively generates intent (imagined movement, attempted speech); the system decodes and controls external devices. This is the mode of most clinical BCIs.
  • Reactive BCI: The system presents external stimuli; the user reacts passively; the system decodes the reaction signal. Typical examples are P300 spellers and SSVEP spellers.
  • Passive BCI: The system monitors the user's cognitive or emotional state without any deliberate intent generation. Typical applications are fatigue detection and attention monitoring.

This distinction is crucial for product design — active BCI requires user training; passive BCI can run invisibly. Apple's AirPods EEG patents take the passive route.


5. Today's "Read / Write" Boundaries

As of early 2026, SOTA boundaries of each capability:

Capability SOTA System / time
Motor imagery decoding (non-invasive) ~80% on 6-class EEGNet baseline
2D cursor control (invasive) 90 bit/min BrainGate + ReFIT
Handwriting BCI 90 chars/min Willett 2021 Nature
Invasive speech BCI 62 WPM, 9.1% WER Willett 2023 Nature
Non-invasive speech decoding 41% sentence recognition Meta Défossez 2023
fMRI → image reconstruction Near-photographic MindEye2 2024
fMRI → video Low fidelity, semantically recognizable MinD-Video 2024
Cortical visual prosthesis Perceivable stable phosphenes Fernández 2021
Intracortical microstimulation (ICMS) 90% tactile detection Flesher 2021

These numbers update every 6–12 months. Subsequent chapters give the detailed technical paths behind each.


6. Shared Concerns with "Human-Like Intelligence" Research

Viewed from the BCI angle, "how to turn the brain into a readable/writable computational object" is itself a human-like-intelligence problem:

  • Representation: Neural population activity lies on a low-dimensional manifold — isomorphic to concepts like "object-centric learning" and "latent-space prediction" in human-like intelligence.
  • Dynamics: Motor cortex as a dynamical system (Churchland-Shenoy) and JEPA prediction in latent space are two sides of the same problem.
  • Learning: Co-adaptation between BCI user and decoder is essentially a meta-learning problem.

These concepts are discussed in depth in Chapter 10 Link to Embodied Intelligence.


7. Logical Chain

  1. BCI = central nervous system × direct pathway × bidirectional information — all three must hold.
  2. The three-dimensional taxonomy (invasiveness / direction / scenario) jointly characterizes a concrete system; any BCI product should be located in this three-axis coordinate system.
  3. Signal scale determines intent granularity — spike-level for fine-grained speech/handwriting; EEG-level for discrete choices.
  4. The 2024-2026 change occurs in the "clinical BCI scenario" dimension: from single trials to multi-company, multi-country commercialization.
  5. Bidirectional BCI is becoming the new frontier — read-write closed-loop is the key from "controlling external objects" to "embodied prosthesis."

References

  • Wolpaw et al. (2002). Brain-computer interfaces for communication and control. Clinical Neurophysiology. — Classic definitional review
  • Zander & Kothe (2011). Towards passive brain-computer interfaces. J. Neural Engineering. — Active/passive taxonomy
  • Hochberg et al. (2006). Neuronal ensemble control of prosthetic devices. Nature. https://www.nature.com/articles/nature04970
  • Willett et al. (2023). A high-performance speech neuroprosthesis. Nature. https://www.nature.com/articles/s41586-023-06377-x
  • Brain-Computer Interface Review (2024). Wiley Brain-X. https://onlinelibrary.wiley.com/doi/full/10.1002/brx2.70024

评论 #