GF ≥ 3σ
loading…
Twelve drifting Fourier modes are the whole visual world. A frozen MaleCNS connectome looks at it through an 8×4 eye, and its 1,314 descending neurons hold the knobs.
shader → eye (8×4 × 6 features) → MaleCNS → descending neurons → Fourier coefficients → shader
A fixed 128-D multiscale Fourier generator defines the luminance world. For the first 8 s DN normalisation is measured with coupling off, then frozen. MaleCNS drives dθ/dt = 0.18 P z(DN) − 0.20 θ through a seeded fixed projection, with θ bounded to ±2. There is no reward or autonomous drift. The primary perceptual state is the 1,771-D retinal luminance vector; θ distance is secondary.
The compound eye looks at the world: 1,771 retinotopic columns (879 left, 892 right), each driving its own lamina L1 and L2 cells, matched by the MaleCNS hex annotation; 1,768 columns have at least one. L3 is left out because MaleCNS v1.0 annotates it in the right eye only, and a one-sided input is an asymmetry a reward-seeking policy would exploit. Light enters with a negative sign because every photoreceptor → L1/L2 synapse in this connectome is inhibitory. The fly acts on its world in 29 ways: yaw and pitch (exact phase ramps), loom (the pattern expands from straight ahead), contrast, the speed of the modes' own drift, and a velocity on each of the 24 mode coefficients. A linear policy from the descending neurons to those actions is trained by reward alone — reward-modulated node perturbation with a one-second eligibility trace — while the connectome stays frozen; with the loop open it neither acts nor learns. The reward is the giant fiber, DNp01 (“GF” in the MaleCNS annotation), central to the visual escape circuit, measured as how far it rises above its own recent mean; it is removed from the policy's input, so the only way to raise it is through what the eye sees. A crossing of 3σ is counted as a takeoff proxy — an event in this rate model, not an observed takeoff. In headless runs a sham learner rewarded by two random DNs crosses as often as the real one, so there is no evidence yet of giant-fiber specific learning. Column geometry, lamina hex coordinates and DN types: MaleCNS v1.0, CC BY 4.0 (Berg et al. 2026, doi:10.1016/j.cell.2026.08.015).
Here the DNs no longer paint the coefficients: the world's texture is fixed and the DNs move the fly through it. A yaw turn is a translation in x, which for Fourier modes is an exact phase ramp c → c·e−i·kx·Δx, so self-motion adds no free parameters. The artifact labels DNs only by side, so steering = mean(right DNs) − mean(left DNs), z-scored against a 20 s baseline, with the convention right > left ⇒ turn right. Left panel of the sweep: open-loop response, Δ(R−L) relative to the paired v = 0 branch. Right: closed-loop yaw; the dashed diagonal is perfect following. The odd part of the response (it flips with the stimulus) is the directional, optomotor-like component; the even part reacts to motion regardless of direction. The sign convention is an assumption: a consistently negative slope would mean the convention, not the fly, is backwards; an even-dominated response means there is no optomotor reflex to find. No null model yet.
One row per mode, brightness = |c|. Stable bands are fixed points; stripes that repeat are cycles; amber ticks mark a perturbation.
Row = mode kicked, column = mode that changed. At each kick the whole state (world, eye, readout, connectome) is snapshotted and a second worker replays four deterministic branches for 3 s: closed loop with and without the kick, and leak-only with and without it. Cell = mean energy over 1–3 s of (closed+kick − closed) − (leak+kick − leak), per unit of kick energy: what the loop did to the kick, with its own drift and the kick’s passive decay removed. Cyan = the loop added energy to that mode, violet = it removed it. This is the operator of connectome plus this arbitrary readout, not of the connectome alone. 0 kicks.