The fly's compass has a built-in brake that no published model includes
The compass is a ring attractor moved by "shifter" neurons that read the activity bump at one position and write it one step over. They are supposed to be one-way conveyors. The wiring shows each shifter gets three times more input from the position it writes to than from the position it reads from. Both left and right shifters do this, so the effect has no direction: it makes the shifters anchor the bump where it is. In simulation the anchor beats the turn signal and rotation collapses. Real flies turn fine, so these synapses cannot be acting the way a naive reading of the wiring says.
Background
- Ring attractor
- ~50 EPG neurons arranged in a ring hold one bump of activity; the bump's position is the fly's heading. Local excitation plus global inhibition keeps exactly one bump alive.
- Shifters
- ~40 PEN neurons, half "left", half "right". Each reads the EPG bump at one position and writes it back to EPGs one step over. When the fly turns, one shifter population gets stronger, so the bump slides. The whole computation depends on shifters being feed-forward: input from position x, output to x+1.
- Where the two contacts live
- The read contact (EPG→PEN) is in a structure called the bridge; the write contact (PEN→EPG) is in the ring. That anatomical split is why every model since 2017 treats read and write as separate, one-way steps.
- Data
- Two synapse-level wiring diagrams: hemibrain (female, 2020, synapses labelled by location) and MaleCNS (male, 2026).
What we found
Each shifter receives EPG input at two places: the expected read position (in the bridge) and, three times more, at its own write position (in the ring). So the EPGs a shifter pushes on push back on it.
| EPG → PEN synapses (hemibrain) | at read position | at write position |
|---|---|---|
| PEN type a | 1,281 | 3,844 |
| PEN type b | 1,512 | 3,645 |
Standard ring-attractor model, constant turn signal, with the write-position feedback added at increasing strength.
| Write-position feedback vs read input | Bump rotation (°/s) |
|---|---|
| 0 (textbook) | 139 |
| 0.5× | 72 |
| 1× | 0 |
| 2.7× (measured) | 0 |
"0" is partly a rate-saturation artefact of the minimal model; the defensible claim is that rotation gain collapses as the anchor strengthens.
Why it matters
- Shifters double as anchors. Normal stalling is two equal pushes cancelling; this is different: both shifter populations pile excitation onto wherever the bump already is, and the anchor feeds on itself. Turning harder also feeds the hold. Models omit this, the largest EPG→PEN pathway; if it works as drawn the compass is stable in darkness but sluggish under rotation.
- Only three ways out, each a different compass: the write-position synapses are functionally weak; they sit on the shifter's axon and modulate its output rather than drive it; or they are cancelled by inhibition (Delta7, ExR4/ExR6) the model lacks.
- Prior work: Turner-Evans et al. 2020 noticed some of these synapses and called them unexpected; Hulse et al. 2021 wrote that "EPG and PEN_a neurons are so densely interconnected within the EB that a network graph of their subnetwork forms a ring" but did not compare it with the bridge route or model it. New here: the 3:1 count against the bridge route, the placement on the write-back tile, conservation across two sexes, and the simulated brake.
What would settle it
- Record a shifter while stimulating EPGs at its read position vs its write position; compare response size.
- At EM level, check whether the write-position contacts sit on shifter dendrites or on their axon terminals.
- Refit a ring-attractor model with the loop; if rotation gain and dark stability cannot both be matched, the loop must be gated or silent.
Wiring: high confidenceConsequence: qualitativeFunction: unknown
Sources: docs/compass-recurrence.md; simulations/eb_epg_pen_loop.py. Turner-Evans 2017, 2020; Green 2017; Hulse 2021; Maimon & Abbott 2026.