A little brain.
A lot to learn.
Meet your newest student. Present a scent. Watch its neurons respond.
CONNECTING TO FLYBRAIN
— independent frozen probes
Awaiting the first measurement
Learning can stabilize; every lesson still runs a fresh measurement
Reward-compartment MBON mean firing rate
Does experience leave a trace?
using the same scent and random input.
Happening in class.
FLYBRAIN’s engine. Our classroom.
This classroom runs our fork of FLYBRAIN: a leaky integrate-and-fire simulation over the male fly CNS connectome. Banana and apple cues become receptor inputs through DoOR odor-response data. Eligible Kenyon-cell → MBON synapses change under the upstream dopamine-gated learning rule.
The graph retains 165,122 traced neurons and 10,228,000 signed connections. Firing rates are simulation outputs. The fly illustration is a render; neural-view points use measured soma positions. The animation replays recorded spikes in 1 ms bins, slowed 100× so they are visible. It repeats the latest result until a new measurement arrives.
Each comparison uses the same scent, random seed, and 40 ms window. Only learned KC → MBON gains differ. These are neural response measurements, not a score for scent-choice intelligence.
We identify reward compartments from incoming anatomical PAM/PPL1 contacts. The upstream fast-weight graph omits dopamine connections, so we use the original unsigned data for this step. Classroom memory is saved separately.
The cue mapping, reward schedule, and simulated dynamics are modeling choices. This experiment does not establish biological learning or internet autonomy. Automatic lessons run while our simulator process is running.
Code © fruitflydev, MIT. Connectome © HHMI Janelia FlyEM, Cambridge Connectomics Group and Google Research, CC BY 4.0. DoOR 2.0 data: Münch & Galizia, CC BY-SA 4.0. Attribution and licenses.