Muskeology
Frontier tech, minus the hype

Neurotech

What we would need to know to build a real brain interface

Reading motor intent works because motor cortex is comparatively well understood. Almost nothing else in the brain is.

Detailed view of a computer processor, showcasing its pins, against a clean white backdrop.
Detailed view of a computer processor, showcasing its pins, against a clean white backdrop. · Photo via Pexels

Existing brain-computer interfaces work by decoding motor commands. Extending them beyond that runs into the fact that neural coding is understood in a small number of places and not elsewhere.

Why motor cortex is tractable

Several properties make it the easiest target.

The output is observable. You can record neural activity while an animal or person moves and correlate the two directly. Ground truth exists.

The mapping is relatively low-dimensional. Movement can be described by a modest number of parameters — direction, speed, force — and populations of neurons encode these in ways that have been characterised.

It is topographically organised. Body parts map to cortical locations in a broadly consistent arrangement, so an electrode's position tells you something about what it will record.

It is on the surface. Accessible to electrodes without traversing other structures.

Sensory cortices share several of these properties, which is why visual and auditory decoding have also progressed.

Why almost nothing else is

No ground truth. For memory, intention, semantic content or emotion, there is no external signal to correlate against. You can ask a person what they were thinking, which is a report rather than a measurement.

Distributed representation. Higher cognitive functions do not appear to be localised in the way movement is. They involve coordinated activity across widely separated regions.

Recording from one area gives a fragment of a process.

Individual variability. Where motor maps are broadly consistent across people, semantic representations appear to vary considerably. Decoders may not transfer between individuals at all.

Context dependence. The same neuron participates in different computations depending on state, task and history. There is no fixed dictionary.

Depth. Structures involved in memory and emotion are deep, requiring penetrating electrodes with correspondingly higher risk.

The scale problem

A human brain contains on the order of eighty-six billion neurons.

Current high-channel-count implants record from hundreds to a few thousand sites, each capturing activity from a small local population.

Scaling channel counts by orders of magnitude runs into power dissipation — electronics in the brain must not heat tissue — data bandwidth, and the physical volume of electrodes displacing tissue.

These are engineering constraints with engineering responses, and they are not trivial ones.

Writing is harder than reading

An asymmetry that is frequently missed.

Electrical stimulation activates neurons non-selectively within a volume — it recruits whatever is nearby, including axons passing through from elsewhere.

Producing a specific, meaningful pattern of activity therefore requires either far finer spatial control than electrical stimulation offers, or knowledge of what pattern to produce, which returns to the coding problem.

Optogenetics offers cell-type-specific control in animals and requires genetic modification, which is a substantial barrier to human use.

Focused ultrasound can modulate deep structures non-invasively and its spatial resolution and mechanism are still being characterised.

Which is why sensory feedback in current systems is crude — conveying touch as a sensation of pressure, not as a specific texture — and why claims about writing information into memory are not close.

What would actually be required

To decode language from thought rather than from attempted articulation: a model of how semantic content is represented that generalises across individuals and contexts.

To write memories: knowledge of how memories are encoded, consolidated and retrieved at a level that permits construction rather than only disruption.

To enhance cognition: a theory of what limits it, and a mechanism to relax that limit without disrupting everything else.

None of these is a matter of more electrodes. They are gaps in basic neuroscience.

What is realistically ahead

Better motor and speech decoding, with more channels, better decoders and longer-lasting hardware.

Richer sensory feedback, improving control and embodiment.

Closed-loop stimulation for neurological and psychiatric conditions, where the target is modulating a network state rather than writing content.

Better understanding from large-scale recording in animals and from the human recordings that clinical implants incidentally provide, which are a genuinely valuable scientific resource.

That is a substantial programme and it is the honest version. The version involving downloading skills is not a longer version of the same road; it is a different road that has not been found.

Ravi Shankaran
Editor, Muskeology

Ravi spent nine years as a powertrain engineer before turning to writing. He is unimpressed by anything that has only ever worked on a stage.

More from Ravi →

Also by Ravi Shankaran

Neurotech

Reading a neurotechnology paper

What to look for in a study, and which sections tell you whether the headline is supported.

Ravi Shankaran··3 min read

Neurotech

Cognitive enhancement claims, examined

Transcranial stimulation, neurofeedback and nootropics all have research literatures, and the effect sizes are smaller than the marketing.

Ravi Shankaran··3 min read

Robotics

Autonomous mobile robots outside the warehouse

Hospitals, hotels, factories and pavements — where wheeled autonomy has spread, and the specific reasons each environment is harder than a warehouse.

Lena Brandt··3 min read