Adam J. Calhoun

Staff Research Scientist · Reality Labs, Meta

I build interfaces between people and machines.

LinkedIn X / Twitter Bluesky Google Scholar Medium GitHub

About

At Meta Reality Labs, I decode the nervous system at the wrist using electromyography, or EMG, so you can control by intention: no keyboard or touchscreen required.

Before that I spent a decade asking why animals do what they do. As a Simons Foundation research fellow at Princeton, I built machine-learning models that uncover an animal's hidden internal states from its behavior alone — tools now used by labs around the world. My PhD in computational neuroscience at UC San Diego showed that worms forage like information theorists. And before that: pure mathematics at St Andrews, solar physics, econometrics, and software at Intel.

Through-line: careful measurement plus good models turns "random" behavior into something you can predict.

Path

2022 –

Staff Research Scientist, Reality Labs at Meta

Developing EMG-based neural interfaces.

2014 – 2021

Research Fellow, Princeton Neuroscience Institute

Machine learning for the neural basis of social behavior; led a team of graduate and undergraduate researchers. Simons Collaboration on the Global Brain Fellow; McKnight–Doupe Fellow.

2007 – 2014

PhD, Neuroscience (Computational), UC San Diego & Salk Institute

Information theory, dopamine, and the science of curiosity-driven search.

2004 – 2007

Master of Mathematics, University of St Andrews

First Class Honours.

2000 – 2003

Intel Corporation

Worked in Intel Architecture Labs on Universal Plug and Play (UPnP); also SQL databases, ASP and JavaScript.

Selected papers

A generic non-invasive neuromotor interface for human–computer interaction

Nature, 2025 · CTRL-labs at Reality Labs

A wristband that reads the faint electrical signals of your muscles and turns them into clicks, gestures and handwriting — and works out of the box for people it has never seen.

Mapping model units to visual neurons reveals population code for social behaviour

Nature, 2024 · Cowley, Calhoun et al., with J. Pillow & M. Murthy

A deep network whose artificial units map one-to-one onto real neurons in a fly's visual system, trained by "knocking out" units just as experiments silence real cells. It reveals how groups of neurons jointly steer courtship.

Unsupervised identification of the internal states that shape natural behavior

Nature Neuroscience, 2019 · with J. Pillow & M. Murthy

A machine-learning method that reads an animal's hidden internal state from behavior alone — predicting a fruit fly's courtship song moment to moment, and revealing which neurons flip it between states.

What is behavior? No seriously, what is it?

iScience, 2023 · with A. El Hady

We surveyed scientists across fields to define "behavior", the thing their fields exist to study. Each person's answer was consistent, but no two fields agreed, and the answers split into at least six distinct kinds.

Neural mechanisms for evaluating environmental variability in C. elegans

Neuron, 2015

Traced how a tiny worm brain uses dopamine to judge whether food is scattered or plentiful — and what looked like random behavior turned out to be memory.

Maximally informative foraging by Caenorhabditis elegans

eLife, 2014

Worms search like information theorists: they explore to learn about their world, not just to eat.

The foraging brain

Current Opinion in Behavioral Sciences, 2015 · with B. Hayden

Why the mathematics of foraging is a unifying lens on decision-making, from worms to monkeys to people.

Quantifying behavior to solve sensorimotor transformations

Current Opinion in Neurobiology, 2017 · with M. Murthy

How measuring behavior with precision lets us crack how brains turn sensation into action.

Writing

Punctuation in novels

Strip away the words and what's left? The punctuation of famous books turns out to be a fingerprint of its author.

Anser anser

Inside one bird

—

PC1 PC2 PC3

Layer one32 units

Layer two32 units

Both hidden layers, fitted to the whole flock's activations so the axes mean the same thing for every bird; the trace is where this one sits along them, over the last sixteen seconds. The shaded band is its alarm, the upright rules are the frames its leave and rejoin gates opened. The outputs are not here — they are read straight off in the panel opposite.

Skeins, learned

One small network, copied into every bird. It sees only what that bird could sense — the slot it should be holding, the bird ahead, a threat, its own pulse — and it decides how to turn, when to call alarm, and when to leave. Press H to hide these controls, S to hide the traces.

Who is flying

Network authority

Cross-fade between the two. Anywhere in the middle both are computed and the wanted heading is mixed, so you can hand the flock over gradually and watch where it wobbles first.

Learning

Steps per frame

Learning rate

The rules label every state the flock actually visits, including states the network drove itself into, so handing over early and leaving training on lets it learn its own mistakes.

Sparsity

Two different things. Weights puts an L1 penalty on the parameters and prunes connections — the network keeps using all its units, but each one listens to fewer inputs. Units penalises the hidden activations instead, pushing each layer toward having only a few units awake at a time, which is what changes the traces next door.

Penalty

Push this too far and the loss will climb — the network is being asked to fly the flock with less. Watching where it breaks is most of the point.

One bird's network

Click any bird to follow it. The left column is what it senses, the right is what it decides, and the pale bar behind each decision is what the rules would have done instead. On the dial, the orange arc is where the leave gate opens and the grey arc where it rejoins.

Drive

Rate

Each bird carries a slow phase of its own, one and a half to three and a half minutes. Nothing outside it can see the phase. When it comes round past zero and nothing is frightening the bird, the leave gate opens — so defection is timed from inside, not rolled for.

Gate window

Visit length

How far round the dial the second window sits. A bird that has left rejoins when its phase reaches it, so this is both how long a visit lasts and what fraction of the flock is away at any moment.

Flock

Skeins

Arm

Gap

Offset

Station keeping

Speed

Turn limit

Wander

Predator

The hawk keeps its own hand-written state machine — it is not one of the flock. Drag on the picture to play the predator yourself.

Between visits

Fear radius

Nerve

How long alarm lingers. The network has no memory except the alarm it sent last frame, so it has to learn this decay itself.

Sky

Hour

Scrub the evening by hand, or let it run. The strip above is the whole cycle: afternoon at the left, last light at the right.

Pace

How long the light takes to cross the whole strip. At zero the sky holds wherever you left it.

At last light

Silhouette

How far the skeins converge to dark as the light fails. At zero they keep their colours all evening, which reads as illustration rather than dusk.

Palette

Plate

Trail

Grain

Wing

Depth

Clip length

Record captures the plate and the traces side by side, drawn fresh at film scale rather than scraped off the panel. MP4 where the browser allows it, WebM otherwise. Saving asks first — the viewer confirms every file the page hands over.

Frame rate

Resolution