A brain-computer interface, or BCI, is a system that reads activity from the brain and turns it into an action in the outside world without going through muscles. Move a cursor by imagining your hand moving. Spell words by attending to letters on a screen. Steer a wheelchair. Control a robotic arm. The idea is simple to state and has taken half a century of neuroscience, signal processing and machine learning to make real.

This article is the map. It explains what a BCI is made of, how brain activity is measured, what signals are actually decoded, what the technology can and cannot do today, and how to start building one yourself. It is also the entry point to the Decoding and BCI track on this site, where each of these pieces gets its own hands-on lesson.

The loop

Every BCI, from a research prototype to an implanted clinical device, is the same closed loop with four stages:

1. Measure. Record a signal that reflects brain activity: electrical potentials from the scalp, the cortical surface, or inside the tissue.
2. Extract. Clean the recording and compute features that carry information: the power in a frequency band, the timing of a voltage deflection, the firing rate of a neuron.
3. Decode. Map those features onto an intention. This is a machine learning problem: a classifier that says left or right, or a regression that outputs a cursor velocity.
4. Act and feed back. Carry out the intention and show the user the result. The user sees the cursor move and adjusts. The brain learns the interface at the same time as the interface learns the brain.

flowchart LR

U((The brain)) --> M[1. Measure]

M --> X[2. Extract features]

X --> D[3. Decode intention]

D --> A[4. Act]

A -- feedback --> U

That last point is the thing people miss. A BCI is not a mind-reading device bolted onto a passive brain. It is a two-way learning system, and much of the practical progress in the field has come from making the feedback loop faster and the user’s job easier.

How the brain is measured

Recording methods trade off three things against each other: how much detail you get, how much of the brain you can see, and how invasive the method is. Nobody has all three.

Most of the BCI research you can do at home uses EEG, because the hardware costs a few hundred dollars and the datasets are public. Most of the headline clinical results use intracortical arrays, because when you can hear individual neurons you can decode far more. The middle of the table is where a lot of current commercial effort sits.

What signals are actually decoded

The raw recording is voltage over time. The decodable information lives in a handful of patterns within it.

Rhythms

EEG has recognisable oscillations, named by frequency band. The bands are conventions, not sharp boundaries, but the usual definitions are:

• Delta, 0.5 to 4 Hz. Dominant in deep sleep.
• Theta, 4 to 8 Hz. Drowsiness, memory tasks, and in animals the hippocampal rhythm that organises spatial memory.
• Alpha, 8 to 13 Hz. Relaxed wakefulness; strongest over visual cortex with the eyes closed.
• Beta, 13 to 30 Hz. Active thinking and, importantly for BCI, motor control.
• Gamma, above 30 Hz. Local cortical processing; hard to see from the scalp, clear in ECoG.

The classic motor-imagery BCI works on one fact about these rhythms: when you move a hand, or imagine moving it, the alpha and beta power over the opposite motor cortex drops. Detect which side dropped, and you know which hand the user imagined. Two classes, one feature, and a working interface.

Evoked responses

Show someone a sequence of flashes and their EEG contains a reliable positive bump about 300 milliseconds after the one they were waiting for. That is the P300, and it powers the oldest spelling interfaces: flash rows and columns of a letter grid, find the row and column that produced a P300, and you have the letter. A related trick, the steady-state visual evoked potential, uses targets flickering at different frequencies; whichever frequency shows up in the EEG is the one the user is looking at.

Spikes and local field potentials

Inside the cortex, the signal is the firing of individual neurons, exactly the spike trains you work with in the first Foundations lesson. Neurons in motor cortex fire in proportion to the direction and speed of an intended movement. Record enough of them, fit a model from firing rates to velocity, and you can drive a cursor or a robotic arm. This is how the best-known clinical demonstrations work.

Decoding is machine learning

Once features are extracted, the decoder is a standard supervised learning problem with unusual data. Linear discriminant analysis and support vector machines still do most of the work in EEG, because the data are small and noisy and simple models generalise better. Kalman filters have been the workhorse for cursor control from spikes for two decades, because they naturally smooth the noisy estimate over time. Deep networks have arrived for the hardest problems, such as decoding attempted speech from cortical recordings, where recent studies have produced text at close to conversational rates for people who had lost the ability to speak.

What makes BCI decoding hard is not the algorithm. It is that the signal changes: electrodes shift, the brain adapts, attention wanders, and a decoder trained this morning is worse by the afternoon. Handling that drift is one of the open engineering problems in the field.

What works today

• Cochlear implants are the most successful neural interface ever built, in use by hundreds of thousands of people. They run in the other direction, computer to brain, but they prove the basic engineering.
• Deep brain stimulation for Parkinson’s disease and other conditions is routine clinical practice, and newer devices adjust stimulation based on the recorded signal, which makes them closed-loop BCIs.
• Intracortical cursor and typing control has allowed people with paralysis to communicate and use computers in research settings for years, and several companies are now running early human trials of implanted devices intended for everyday use.
• Speech decoding from cortical recordings moved from proof of concept to conversational speed in the last few years.
• Consumer EEG headsets can reliably detect a few coarse states such as eyes closed, drowsiness or attention to a flickering target. They cannot read thoughts, and the ones that claim to are selling the band names above, not neuroscience.

Why this matters for the rest of this site

A brain-computer interface is where every track on this site converges. You need to understand what the recording is measuring, which is the Signals track. You need to handle real data in standard formats, which is the Data track. You need a model of what neurons are doing, which is the Models track. And you need to build a decoder that works on a live stream and keeps working, which is the Decoding and BCI track itself. If you follow the path from the beginning, a working motor-imagery BCI on public EEG data is the destination.

How to start

You do not need hardware. Everything below is free.

• Data. The PhysioNet EEG Motor Movement/Imagery dataset has over a hundred subjects doing exactly the left-versus-right task described above. The BNCI Horizon 2020 collection has dozens more BCI datasets in one place.
• Tools. MNE-Python for loading, filtering and visualising EEG and MEG. scikit-learn for the decoder. MOABB for benchmarking BCI algorithms across datasets without rewriting the loading code each time.
• Hardware, later. When you want to close the loop on your own head, open-source boards such as OpenBCI cost a few hundred dollars and stream directly into Python.
• Reading. Wolpaw and Wolpaw’s Brain-Computer Interfaces: Principles and Practice is the standard reference and covers everything above in depth.

The author declares no conflict of interest.