Methodology

Two separate experiments have lived under TimeFly. The interactive page runs a connectome simulation, and everything it draws is described in part one. The Vision Lab, an earlier demonstration of a different, pretrained visual-system model, is no longer shown on the page; its methods stay documented in part two. Neither experiment is evidence about the other, and neither measures an animal.

What the page draws

The brain panel places every published neuron at its annotated soma position from the MaleCNS release, projected from the front, and flashes it when the simulation reports a threshold crossing. Sensory neurons whose cell bodies lie outside the imaged nervous system carry no soma coordinate; they are drawn at the mean position of their synaptic partners in the published graph. That is a drawing choice for a picture of activity, not an anatomical claim, and the map file records how many neurons it applied to. The eye panel draws one hexagon per measured optic-lobe column at its lattice coordinate, filled with the luminance sampled under that column and tinted by the signed lamina contrast. Under the chase rule the light moves to a new seeded position each time the fly reaches it, so a film plays to its end; the path and the network state are never reset by a catch.

Part one: the connectome experience

Engine version timefly-engine-1.0.0.

What is measured

The wiring diagram. MaleCNS v1.0, released by Janelia Research Campus under CC BY 4.0. The full release annotates 211,577 bodies. The graph published here holds 174,252 neurons, 6,133,500 connections and 88,354,504 synapses, which is a selected subgraph: it keeps only connections of at least five synapses between annotated bodies, and drops connections whose presynaptic transmitter could not be predicted. Every filtering rule and the count it removed are recorded in the published bundle manifest. This is a measured anatomical reconstruction with annotations. It is not a functioning brain, and it does not come with dynamics.

What is modeled

The dynamics. Every neuron uses one leaky integrate-and-fire model with the same parameters: resting and reset potential -52 mV, threshold -45 mV, membrane time constant 20 ms, refractory period 2.2 ms and a uniform conduction delay of 1.8 ms. Those values follow Shiu, Sterne and colleagues (Nature, 2024). At the frozen timestep of 1 ms the refractory period and the delay each round to two steps, so both are 2.0 ms in practice.

One thing departs from that published parameterisation, deliberately. Their model applies a fixed 0.275 mV step per synapse. Against this connectome that does not work at either end: in-degree spans four orders of magnitude, so a fixed step either drives the hub neurons to saturation or never reaches the sparse ones, and both failures were measured. Synaptic drive is therefore normalised by each neuron's own measured total input weight, with a single gain of 2.6: a neuron fires when roughly 0.38 of its measured input is active together, whether it has ten inputs or ten thousand. The synapse counts used are measured. The normalisation is a modeling choice.

Connection sign comes from the released transmitter predictions: acetylcholine is treated as excitatory, and GABA, glutamate and histamine as inhibitory. Neuromodulators are not modeled. There are no gap junctions, no synaptic time constant and no plasticity. Nothing learns.

What the model sees

The film is treated as a flat visual world the fly moves across, not a reconstruction of the filmed scene. Input enters at the lamina monopolar cells L1, L2, L3 and L5, because those are the cells this release gives measured retinotopic column coordinates for on both optic lobes. The photoreceptor bodies themselves carry no column assignment here, so using them would have meant inventing positions.

Each column looks along its own direction at its own distance and reads the frame the visitor is watching, including the beacon. Luminance is low-passed at 10 ms, adapted over 250 ms, and expressed as contrast against that adapted background. L1 carries the increment half and L2 the decrement half, matching the split that feeds the fly's motion detector. A completely static scene therefore produces almost no drive once adaptation settles, which is why movement, not brightness, is what the network responds to. Sampling geometry is resolved every 4 steps, a 250 Hz hold on the receptor input.

What is engineered

The game interface. Steering reads a laterality index across a named descending population: the anterior-dorsal cluster DNa01 to DNa16, sixteen neurons per side, which contains the published steering neurons DNa01 and DNa02. Three candidate readouts were registered before any footage was used: the DNa02 pair alone, the DNa01 and DNa02 pairs, and the whole cluster. Calibration on synthetic stimuli selected the cluster, because one or two neurons per side do not produce a usable graded signal in this model. That choice was frozen before any archival clip was evaluated.

The index is (right minus left) divided by their sum, so it does not depend on how hard the network happens to be driven; an absolute voltage threshold would have been a hidden tuning knob. Below a combined activity floor of 0.00002 the index is treated as no signal and the fly flies straight. A turn begins past an index of 0.06 and scales at 190 degrees per second per unit of index, capped at 300. Forward motion is an engineered tonic drive of 0.155 frame widths per second, suppressed by the published freezing neuron DNp09 and briefly raised by the giant-fibre readout DNp01.

The beacon is an artificial stimulus drawn into the film and shown to the model. The portal radius is 0.055 frame widths. A journey runs to at most 3 scenes. The fly avatar's banking and wing motion are drawing, not biomechanics: there is no gait, no aerodynamics and no muscle model.

What was verified, and what was not

Calibration used synthetic stimuli only: a uniform field, a static grating, gratings drifting in four directions, a looming disc, and a beacon on each side with a matched no-beacon control. Recorded results: a uniform field and a static grating produce no spikes at all; the steering index sits at zero with no beacon, goes negative with a beacon on one side and positive with it on the other. Zeroing the descending readouts leaves the fly flying perfectly straight. Swapping left for right mirrors the turn. The same seed reproduces the same path exactly. Changing the film condition changes the path.

What is not established: that this model tracks anything, predicts motion, recognises a car or a person, or resembles what a real fly would do in front of a projector. The fly reaching the beacon is not the point and is not guaranteed. A run that fails is the actual result of the recorded simulation, and it is shown as such.

Part two: the Vision Lab

Methods version 0.1.0. Historical-film transfer is a new experiment. The upstream paper does not establish performance on Stockfilm footage.

Input and time

Each excerpt uses a documented square crop, aspect-preserving bilinear resize, and grayscale conversion (0.2126 R + 0.7152 G + 0.0722 B) / 255. This uses decoded RGB values without gamma inversion; it is not calibrated biological luminance. Flyvis BoxEye applies its mean box filter to the receptor lattice specified by the checkpoint.

Original presentation timestamps determine which decoded observation is held for each 0.01-second solver interval. Reduced temporal sampling retains every second decoded frame at its original timestamp; it preserves elapsed time. Lower contrast applies 0.5 + 0.5 (x - 0.5), clipped to [0, 1], before receptor rendering. No per-frame renormalization is applied.

Model and initialization

Flyvis 1.2.0, official flow model 0000/000, checkpoint 00000, is frozen in evaluation mode. Each variant begins independently with one second of gray input and a half-second fade to its first frame. Warmup is excluded from the public timeline. The recorded response is the state after each solver update.

What the response means

The selected population is T4a, indexed from the connectome. Values are signed modeled node activity in arbitrary model units. They are neither spikes nor recordings from an animal. One reference min/max scale is fixed across a clip's variants, with clipping disclosed; amplitudes across clips are not directly comparable.

Evaluation status

The motion-direction task requires reviewed translational regions, known timing and provenance groups. No reviewed scored subset is published. A conventional phase-correlation baseline is separate from Flyvis. Flyvis task scoring is unavailable until a compatible decoder and coordinate conversion are verified. No superiority claim is made.

Sources

MaleCNS connectome v1.0 · Google Research announcement · Lappalainen et al., Nature (2024) · Official Flyvis implementation

Connectome data is used under CC BY 4.0 with attribution to Janelia Research Campus. No endorsement by Janelia, Google Research, TuragaLab or any other party is implied.

Footage rights and separately scoped dataset agreements govern use of Stockfilm material. A footage licence does not grant model training, redistribution, connectome research or dataset rights.