We copied the connections between 499 reconstructed fruit-fly neurons into software and fed the resulting network human voices. A small output layer learned to predict labels listeners had assigned to those recordings. The fly circuit stayed fixed throughout training.
This was a computational experiment with an artificial input pathway and simplified neuron dynamics. No living fly learned to recognize human feelings.
Disclosure: I'm a cofounder of Oruk, where this experiment was conducted. This adaptation was prepared by an AI agent from our published research article and checked against its reported methods and results. The experiment has not been peer reviewed.
The MaleCNS reconstruction supplies neuron shapes and connection tables. We selected a strongly connected group: every neuron can reach every other through some directed path. It contains 15,865 connections supported by 867,344 synaptic contacts.
More contacts produce stronger model weights. Neurotransmitter annotations supply positive or negative signs, followed by global spectral scaling. These are modeling assumptions; the wiring diagram alone cannot recreate each cell's physiology.
That recurrent wiring gives the network a short memory. Every 10 milliseconds, a new frame of audio enters through a fixed random projection. Each neuron combines its neighbors' previous activity with the incoming sound and retains some of its previous state:
x[t] = (1 - alpha) * x[t-1]
+ alpha * tanh(W @ x[t-1] + W_in @ u[t])
Here, u[t] contains 32 log-mel audio bands, x[t] contains 499 neuron states, and both weight matrices stay fixed. The interactive model uses alpha = 0.15 and a target recurrent spectral radius of 0.9.
This is reservoir computing. Suárez and colleagues developed conn2res to explore biological connectomes this way; Costi and colleagues have used fly connectivity for time-series prediction. We applied the approach to listener judgments of speech.
We resample audio to 8 kHz, use 25 ms analysis windows, and normalize features using training frames. Four successive temporal averages turn the circuit's activity into 1,996 features. Another 128 features summarize the audio directly. Both branches enter the learned ridge readout:
features = concatenate(pooled_neuron_states, pooled_audio)
scores = features @ beta + bias
Only beta and bias are learned. The output has 15 emotion labels and 16 speaking styles. The direct audio branch matters: useful predictions can come from acoustic features without passing through the fly circuit.
Training used 16,995 clips, validation 2,239, and testing 2,022. Known speakers stay within one split; when speaker identity is unreliable, the entire source corpus stays together. Exact duplicate audio does not cross splits. Most recordings have one listener rating, which limits what their labels can establish.
The decisive comparison was scrambled wiring. We shuffled connection destinations while preserving each source neuron's outgoing count, signs and weight values before rescaling, then selected settings and trained a separate readout. We also tested audio-only and disconnected-circuit controls.
| Model | Test mean average precision |
|---|---|
| Fly wiring | 16.84% |
| Scrambled wiring | 16.88% |
| Constant scores | 9.71% |
Mean average precision measures ranking across the 31 labels; it is not classification accuracy. The fly-minus-scrambled difference was −0.04 percentage points, with a 95% interval of −0.16 to +0.07. The comparison averages three seeds; intervals use paired bootstrap resamples of 45 held-out speaker/corpus groups. These results provide no evidence that the fly's particular wiring helped on this split.
The interactive article also lets you silence neurons while holding the fitted readout fixed. Removing the 50 highest-ranked neurons drops that model's test mAP from 16.91% to 10.83%. Its starting score differs from the table because this intervention uses one model, rather than the three-seed mean.
That readout depends on its circuit's features. Destroying those features after training can hurt even when a different circuit, trained separately, works equally well. The intervention and scrambled control therefore answer different questions.
You can rotate the reconstructed neurons, compare six licensed demonstration recordings and rerun the intervention in the original article. The clips differ in speakers and words as well as listener labels; colored activity is not evidence of biological “emotion cells.” Those interactive figures cannot run inside this DEV post.
The fitted demonstration model and six excerpts are public. The full speech-labeling dataset and listener records remain private, so this is not a fully reproducible public benchmark. The fly circuit is an exploratory model, separate from Oruk's production speech API.
Top comments (0)