What’s real here, and what isn’t

Measured (MaleCNS v1.0, Janelia · Cambridge · MRC LMB · Google, CC BY 4.0): which neurons exist (165,122 traced), where each soma sits, the shape of each arbor, every connection of 10+ synapses between them (2.7M), each cell’s predicted transmitter, the eye’s column map, and which cells are male-specific, dimorphic or fru/dsx-expressing.

Authored: how activity moves. The wiring has no voltages or timing, so a simple rate model runs on it: inputs weighted by synapse count and signed by transmitter, then decay and fatigue. Every number in that model was picked, not measured, and its milliseconds are model steps (1/30 s), not the real fly’s. Your camera drives the eye by how much each patch changes; the eye’s orientation onto your image is approximate.

What the brain decides, and what we do: in Say hello, whether the fly flinches is read off its escape neurons (giant fiber DNp01, and DNp02, DNp04, DNp11), and which way it jumps comes from which optic lobe lit up more. Walking, the jump itself, and where the swatter comes from are animation.

Fruit Fly Airlines: the game world is rendered from the fly’s head into its eye columns every frame. When its brain flies, it uses a collision-avoidance reflex. We mapped where in its view each of its looming detectors (LC4, LPLC2, LPLC1, LC16) looks, the way neuroscientists map receptive fields: by showing the model’s eye small shadows all over and recording which cells respond. Each moment it compares how many are firing in the lower half of its view (a building rushing in) with the upper half (clouds): more below, it jumps; more above, it holds. Its escape neurons also make it jump, but only away from the threat. The balance is averaged over a quarter of a second, with a little random noise, as real neurons have, so no two flights are the same. That rule and its threshold are ours; which cells fire, and when, is the model on the real wiring. Its view is contrast-stretched first, as the cells behind a fly’s photoreceptors do. Between jumps it glides, then waits 0.7 s before it can jump again. The physics, the course, the landmarks, the scoring and how weather pushes the fly are ours. Each city’s time of day is real: the sun sits where it is in that city’s sky right now (worked out from its latitude, longitude and the date), so it can be night in London while it’s morning in Dallas, and each city shows its local time. Each city’s weather is its real weather right now (from Open-Meteo), except that dust and smog stay on clear days, and tornadoes appear only in tornado country during real rain or storms (a tornado doesn’t kill it: it whirls it up and spits it out past the city). Which cities kill its brain follows from how each landmark’s shape and contrast drive its eyes. Its neurons drive the drums: optic-lobe activity sets the hi-hats, looming the snare, and escape spikes the boom.

Watch TV: each video frame is turned into how much each of the fly’s eye columns sees change, exactly as with your camera, and the soundtrack’s loudness drives its antennae. It flies off only when its escape neurons fire. Two rules are ours: it comes back after 4 seconds, and after landing it needs 3 seconds to settle before it can flee again (real flies’ escape reflex also tires with repetition). It reacts to motion and sudden change, not to what a video means. YouTube videos can only be seen if you share the tab; nothing leaves your browser.

The body is NeuroMechFly v2 (NeLy lab, EPFL, Apache-2.0): 3D meshes from a micro-CT scan of a real fruit fly, in its real joint hierarchy, walking with joint angles recorded from real flies. It is a different fly (female) from the brain (male), so the nervous system is scaled by 0.8 to fit inside it. The glass look, bristles and wing beat are ours. Regions are assigned by cell-type name; the light running along active arbors is decoration.

Nobody knows what the fly is thinking. Source and method.