The Fly's Compass
Inside a fruit fly's brain, a ring of neurons keeps track of which way the fly is facing. One patch of activity sits on the ring, and where it sits is the fly's heading. The card asked whether that patch is kept by the circuit or only carried while something outside holds it.
A compass needs something outside it, the Earth's magnetic field, to point anywhere. The fly's heading circuit mostly does not. It takes in two signals, how fast the fly is turning and where familiar landmarks are, and it uses them to move its patch of activity around the ring. Take the landmarks away by switching off the lights, and the patch stays where it was. Something inside the circuit is holding it there, and the experiment described below shows what.
The findings about the fly are drawn from Kim, Rouault, Druckmann and Jayaraman (2017), with the date read; the paper's own words are in quotation marks. The panel is a standard textbook model of this kind of circuit, built to the wiring the paper reports, and every number from it was measured from the code before the page was written. It is a teaching model and was not fitted to the fly's data.
A bump on a ring
The neurons in question sit in a doughnut-shaped structure in the middle of the fly's brain called the ellipsoid body. The authors imaged them in flies flying while tethered in place. At any moment only a few neighbouring neurons were strongly active: a single bump of activity. As the fly turned, the bump moved around the ring. In the paper's words, "The activity bump closely tracked the fly's heading in flight (Fig. 1K) and persisted in darkness (Fig. 1H)."
The last words matter most. In the dark the fly has no landmark to see, yet the bump stayed. It did not stay perfectly still: "the bump drifted gradually" around the ring. A heading kept with no outside reference slowly goes wrong, the way a person counting paces in a dark room slowly loses track of where they are.
the fly turns → a turning signal slides the bump around the ring → landmarks, when the fly can see them, nudge the bump into line → the ring's own wiring holds the bump in place between nudges
The test: make a second bump
To find out what holds the bump, the authors switched on the neurons at a chosen spot on the ring with light, using a light-sensitive protein put into those neurons (a method called optogenetics), while the original bump sat somewhere else. A ring that simply passed its inputs through would then show two bumps. The fly's ring did not. "As the new bump formed, activity at the previous location began to decline and eventually disappeared". And when the stimulating light went off, "the amplitude of the artificially created bump settled at levels typically evoked by sensory stimuli and did not disappear." The ring held the new bump as if it had always been there.
Two properties of the wiring explain this, and the paper names them: "A network with local excitation and global inhibition enforces this unique and persistent heading representation." Neighbouring neurons excite one another, so an active patch keeps itself going. Every neuron inhibits every other one, so only one patch can survive. The authors measured the effective wiring and found it "consistent with ring attractor models characterized by narrow local excitation and flat long-range inhibition".
The ring, run here
Sixty-four model neurons on a ring, wired as the paper describes: each one excites its near neighbours and inhibits all the others. A neuron with no input from its neighbours falls silent, so any bump that lasts is made by that wiring. The panel runs in real time.
How to read the picture. On the left, the fly in its arena, seen from above, with the landmark at the top of the wall. On the right, the ring of neurons: bright wedges are active, dark ones are silent. The white tick outside the ring is the fly's true heading. The gold tick inside it is where the bump is. When the circuit is doing its job, the two ticks line up. Click anywhere on the ring to stimulate the neurons there for one second, as the paper did.
How to read the graph. The fly's heading (white) and the bump's position (gold) over the last thirty seconds. A gap in the gold line means there was no single bump to follow. Blue ticks mark stimulation, pale yellow the lights switched, red a scramble, white the wiring cut or restored.
Measured from the panel
| what was done | what happened |
|---|---|
| activity scrambled at random across the ring (50 runs) | one bump in 35 runs; in the other 15 the activity died out and the ring fell silent |
| ring silent, lights off, 120 s | stayed silent in all 20 runs: the noise alone never starts a bump |
| ring silent, lights on | the landmark started a bump at the fly's heading in all 20 runs, within 0.1 s |
| that bump, lights then switched off, 60 s | still there in all 20 runs |
| fly still, lights off, 30 s | one bump at the end of every run; it drifted 8° on average (root mean square), 15° at most |
| fly still, lights off, 60 s | drift of 10° on average, 19° at most |
| fly still, lights on, 30 s | bump within 2° of the true heading on average |
| fly turning steadily at 30, 90 and 180°/s (5 runs each) | the bump moved at 0.99, 1.00 and 0.99 times the fly's turning speed |
| fly wandering at random (typical turning speed 57°/s), 60 s, lights off | bump 9° from the true heading on average, 16° at most |
| fly wandering at random (typical turning speed 57°/s), 60 s, lights on | bump 2° from the true heading on average, 5° at most |
| stimulated 180° or 135° from the bump for 1 s (10 runs each) | the old bump died out and the new one stayed, still there 3 s after the light went off, in every run |
| stimulated 90° or 45° from the bump for 1 s (10 runs each) | one bump every time, but dragged only part of the way: to 70° (from 90) and to 32° (from 45) |
| wiring cut, lights off (10 runs) | the bump gone in 0.06 to 0.07 s |
| wiring cut, lights on, fly wandering, 10 s (10 runs) | a faint patch where the landmark falls, peak activity 0.12 to 0.19 against 1.0 for an intact bump; gone within a second of the lights going off in every run |
| wiring cut for 2 s in the dark, then restored, 60 s | no bump in any of 20 runs: the ring stayed silent |
Three results are worth reading twice. With the lights off, the bump holds for a minute with nothing outside the circuit telling it where to be, and it drifts. With the wiring cut, it is gone within a tenth of a second, and switching the wiring back on in the dark brings nothing back: the ring stays silent. The knowledge of the heading lived in the bump, and the bump lived in the wiring. And the landmark can start a bump on a silent ring, but once started, the bump no longer needs it: switch the lights off and it stays.
One row does not match the fly. The paper reports that a nearby stimulated bump moved all the way to the stimulated spot. In this model, stimulation within about 90° of the bump drags it only part of the way. The model's neurons saturate, so the stimulated neurons cannot outshout the neighbours of the old bump. The model is simpler than the fly here, and the page does not rest anything on that row.
Layer 2 of 2 · the framework's reading
The wrapped reading
The Shape of the Basin
The attractor is the ring of one-bump states: one bump of the usual size, at any heading. The basin is the set of starting patterns of activity that grow into such a bump. In the dark the model has a second attractor, silence, and every start that does not reach a bump ends there. The numbers below were measured with the lights off, from starts shaped as a single patch of activity of a given height and spread, ten runs each, recorded in full in the sweep record.
The edge of this basin is a band of odds. Every neuron gets a little noise, so the same start can catch in one run and die in the next. Counts like "5 of 10" are return rates, and near the edge there is a band where both outcomes happen.
What decides membership is concentration. A patch spread over 29 to 46° of the ring catches from a peak of about 0.12 (5 of 10) and does so every time from 0.2. A narrower patch needs more height: spread over 11° it needs 0.3, over 6° it needs 0.5, and over 3° only the full 1.0 caught. A wider patch, over 69°, needs about 0.2. Activity spread evenly around the whole ring never caught at any level up to 1.0, because the inhibition every neuron sends to every other cancels it. What matters is how much activity sits together in one place. Activity scrambled at random caught in 35 runs of 50 and died in the other 15.
The other side of the edge is permanent in the dark. A ring that fell silent stayed silent in all 20 runs over 120 s. Cutting the wiring throws the ring across the edge within 0.07 s, and restoring the wiring in the dark does not bring it back (0 of 20).
Along the ring there are no walls at all. Push the bump to a new heading, by stimulating 135 or 180° away, and one bump of the usual size comes back at the new place in every run; nothing pulls it back to where it was. With the lights off it drifts, 10° on average over 60 s. With the lights on, the landmark adds a pull toward the fly's true heading and holds the bump within 2° of it. That pull is a basin for the position, and it is supplied from outside: it lasts exactly as long as the light does. The basin for the bump's existence belongs to the circuit, and the basin for its position is handed in.
The Morph of the Basin
The strength of the excitation between neighbouring neurons is the dial that reshapes this basin. The panel does not expose it; it was turned offline, with the rest of the model left as shipped (40 is the panel's setting). Changing it rewrites the rules, so each row is a different model, each with its own attractors. Ten runs per cell; "smallest start that catches" is the lowest peak, with the patch spread over 29°, that grew into a bump in at least 5 of 10 runs.
| excitation | full bump holds 10 s | smallest start that catches | silent ring stays silent, 60 s |
|---|---|---|---|
| 50 | 10 of 10 | 0.08 | 3 of 10 |
| 45 | 10 of 10 | 0.10 | 10 of 10 |
| 40 | 10 of 10 | 0.15 | 10 of 10 |
| 35 | 10 of 10 | 0.20 | 10 of 10 |
| 30 | 10 of 10 | 0.40 | 10 of 10 |
| 28 | 10 of 10 | 0.50 | 10 of 10 |
| 27 | 10 of 10 | 0.70 | 10 of 10 |
| 26 | 0 of 10 | none | 10 of 10 |
Weaken the excitation and the basin shrinks. The start needed to catch rises from 0.15 to 0.7 as the excitation falls from 40 to 27. Between 27 and 26 the bump attractor itself disappears: a full bump dies out, and every start falls silent.
Strengthen it and the basin grows until silence loses its own. At 50 a start of 0.08 catches, and the noise alone starts bumps on a silent ring: silence held for 60 s in 3 runs of 10, and in none over 120 s. At that point every state eventually ends in a bump and there is no edge left to describe.
Noise moves the same edge. With the excitation at 40, a silent ring stayed silent for 120 s in every run up to a noise level of 0.26, in 9 of 10 at 0.28, 5 of 10 at 0.30 and none at 0.35. More noise also means more drift in the dark. The panel's noise, 0.24, was chosen to sit below that boundary. An earlier build of this page ran at 0.30 with thirty-two neurons and had no edge at all: noise started a bump from silence in 40 runs of 40.
What the morph can and cannot say here. Every basin above is drawn inside a model whose neurons, wiring and running costs are supplied: no maintenance cost is represented anywhere in it, and in the fly all of it is made and paid for by the animal. The morph shows which dials make a bump possible and how much of a start it needs. It says nothing about whether the circuit could hold its shape while paying for itself. The existence of a silent state in the fly's own circuit is not claimed.
What would show this reading wrong
A bump that persisted in the dark only because some other part of the brain kept re-supplying it at the same place, which would make the ring a carrier of a position held elsewhere.
Two bumps holding side by side, stably, under conditions where local excitation and global inhibition predict that only one survives.
A bump that survived with the ring's own excitatory connections blocked, which would mean the wiring is not what holds it.
Honest limits
The paper was read in passages, through a web fetch of a course copy of the article, not in full, and its supplementary material was not read.
The model has sixty-four neurons, one kind of neuron, and chosen strengths. The fly's circuit has several interacting neuron types, and the turning signal reaches the ring through some of them; the model folds that into one term. The model's drift in the dark comes from noise, at a level chosen to make the drift visible while keeping a silent ring silent, not fitted to the fly. An earlier build with thirty-two neurons could not do both: noise low enough to keep the ring silent also locked the bump in place between neighbouring neurons, an artifact of the grid.
The paper studied flying flies. Nothing on this page speaks to other insects or to the heading cells of mammals.
Sources
Kim, S. S., Rouault, H., Druckmann, S., & Jayaraman, V. (2017). Ring attractor dynamics in the Drosophila central brain. Science 356(6340), 849–853. Source of every quoted finding on this page: the bump tracking heading and persisting in darkness, its gradual drift, the optogenetic test, and the local-excitation, global-inhibition wiring. doi.org/10.1126/science.aal4835
Related on this site: Life & Physiology; Homeostasis, another loop inside a larger one; and the Circadian Clock, a rhythm that also keeps going in the dark.