Neuron positions · Drag to rotate
Robofly plays four recorded simulations of a fruit-fly neural network responding to visual input. The video and brain activity stay in sync. All simulations were computed in advance.
The simulation uses the full retained MaleCNS connectome: 166,700 neurons and 25,582,938 directed connections, based on mapped fruit-fly anatomy. Its connection weights stay fixed.
This experiment combines the simulated brain with a small trained model to follow a moving ball.
That model, called a readout, was trained on neural activity produced by the simulation while it viewed synthetic balls. It learned to translate those responses into movement of the tracking halo.
At each step, a 128 × 128 image around the halo’s previous position enters the simulated eyes. The network responds, and the readout uses that activity to move the halo. We trained the readout; the connectome’s weights were not changed.
Each image runs for 200 simulated milliseconds. Neural activity resets between images, but the halo’s position carries forward.
The combined system tracks a simple synthetic ball. We haven’t established whether the fly wiring helps it track better than a simpler system, or whether it can reliably track objects in real footage.
These experiments compare responses to light on either side of the image and to moving stripes. They use the full image, without a tracking readout.
Neural activity carries forward between images. Each image runs for 100 simulated milliseconds after an initial 500 milliseconds of black input. We compare the result with a matching simulation that receives black input throughout.
The 124,295 dots show measured positions of neuron cell bodies. They don’t form a solid brain surface, so some groups sit outside the denser central shape. Neurons without a displayed position, including nerve-cord neurons, are still part of the simulation.
Lavender shows a change in simulated spike count relative to a reference response. That change can be an increase or a decrease. Dim dots show neurons whose spike counts did not change.
Round & round uses a black retinal image with constant background input to the lamina, an early visual-processing layer, as its reference. The other recordings use the matching black-input simulation described above.
The glow makes changes easier to see. Its size doesn’t represent neuron size, and its brightness isn’t a confidence score.
This is a simulation built from mapped fly wiring. It doesn’t establish that the network sees or understands the world as a living fly does. The tracking readout has not been validated as a general-purpose object tracker.
Choose a short video. We’ll process its first 12 seconds at five samples per second using the full simulation on this computer.
A bright object against a dark, uncluttered background is closest to the training images. Other scenes still produce neural activity, but the halo may drift.
