Intelligence Mapping via Neuroplasticity in the Fruit-Fly Connectome

What does a small fruit-fly brain teach us about big ones?

Venue: Alameda County Science & Engineering Fair, Biotechnology — 4th Place, HS-BCOM-347
Connectome: FlyWire — access won by written request
Tools: NetworkX for the connectome graph; Brian2 for the spiking simulation
Neuron model: Leaky integrate-and-fire neurons; spike-timing-dependent plasticity at the synapses; dopamine-modulated plasticity for reinforcement
Year: 2025

186,000neurons in the fly brain
120neurons in my simulation
5cell types modelled
2 × 2odour cases: attract or repel, weak or strong

The question

A human brain has 86 billion neurons; a large language model has 500 billion parameters. Both are far too big to trace. A fruit fly has 186,000 neurons, and its wiring has been mapped in full. So: what does a small brain teach us about big ones? Concretely, three things to simulate — how the fly tells an attractive odor from an aversive one, how it tells a strong odor from a weak one, and what spike-timing-dependent and dopamine-modulated plasticity each contribute.

Building the circuit

Five steps. Extract the olfactory network from the FlyWire connectome with NetworkX — which neuron types, how many, and how they connect. Rebuild it in Brian2 as 120 simulated neurons, leaky integrate-and-fire cells joined by plastic synapses. Model the odour stimulus. Add reinforcement-modulated learning. Simulate, and read the raster plots and weight curves.

The simulation is deliberately small — small enough to trace every connection, which is the whole reason to start with a fly. Its five cell types and their wiring come from the real map. These are the counts I read off it, and the proportions the model keeps:

Cell type	In the connectome	Share	Role
Olfactory sensory neurons (OSN)	~1,400	32%	Detect odor in the antennae
Projection neurons (PN)	~400	9%	Relay from the antennal lobe to the mushroom body
Local interneurons (LN)	~200	4.5%	Local processing within the antennal lobe
Kenyon cells (KC)	~2,200	50%	High-dimensional odor representations
Output neurons (ON)	~200	4.5%	Project toward the motor circuits

And the connection densities between them, again read off the map rather than assumed:

Pathway	Connections in the map	Density used
OSN → PN	165,000	30%
OSN → LN	138,000	50%
LN → PN	34,000	40%
LN → KC	43,000	10%
KC → ON	130,000	30%

Results

Attraction vs. aversion. The projection neurons, local interneurons, and Kenyon cells fire in nearly the same pattern for both odors. Only the output neurons differ — and that difference is the decision. Activity also persists after the stimulus ends. The same network distinguishes conflicting concepts.

Strong vs. weak. A weak and a strong version of the same odor produce the same output activity, for attraction and for aversion alike. The fly can tell which odor is present, not how strong it is.

Reward and punishment. Spike-timing-dependent plasticity on its own only strengthens synapses. Add a dopamine signal — positive, then negative — and the same pathway specializes for opposite meanings: the average synaptic weight climbed to its ceiling by about trial 20, fell back to baseline by about trial 55, and climbed again by about trial 70.

What it means

Through neuroplasticity, the same network can be transformed to learn conflicting concepts. Artificial neurons store multiple concepts too — which is what makes them hard to study — so the tools built for biological neurons may work on artificial ones.
Smaller networks do not perceive gradations of a concept. A neuron-constrained brain learns black versus white; larger networks are needed to see the shades between.
Biological and artificial neural networks share learning characteristics and can be studied with similar tools and methods.

What’s next

Scale the model toward the real network to see how the learning principles change with size; let the output neurons’ feedback reach every neuron on the pathway, not only the Kenyon cells; and use what works for overlapping concepts in biological networks as a guide toward understandable AI.

Rotating 3D morphology of an olfactory sensory neuron from the fruit-fly connectome — Fig. 1. Olfactory sensory neuron.

Rotating 3D morphology of a projection neuron from the fruit-fly connectome — Fig. 2. Projection neuron.

Rotating 3D morphology of a local interneuron from the fruit-fly connectome — Fig. 3. Local interneuron.

Rotating 3D morphology of a Kenyon cell from the fruit-fly connectome — Fig. 4. Kenyon cell.

Rotating 3D morphology of a mushroom-body output neuron from the fruit-fly connectome — Fig. 5. Output neuron.

Population firing rates over time and mean response per cell type for a weak attractive odor — Fig. 6. Weak attractive odor: firing over time, and mean response during vs. after the stimulus, per cell type.

Population firing rates over time and mean response per cell type for a strong attractive odor — Fig. 7. Strong attractive odor.

Population firing rates over time and mean response per cell type for a weak aversive odor — Fig. 8. Weak aversive odor.

Population firing rates over time and mean response per cell type for a strong aversive odor — Fig. 9. Strong aversive odor.

Histogram of synaptic weights before and after learning, and the learning curve under positive then negative reinforcement — Fig. 10. Synaptic weights before and after learning, and the learning curve under positive, then negative, reinforcement.

Preview of the ACSEF 2025 quad chart, Intelligence Mapping via Neuroplasticity in the Fruit Fly Connectome — Fig. 11. The ACSEF 2025 quad chart (HS-BCOM-347). Click to open at full size.