Following the publication of the digitized Drosophila brain (the FlyWire connectome project), the internet was flooded with demonstrations allegedly showing a virtual fly driving a car or playing classic 3D shooters. However, behind these flashy videos often lies simple sensory-motor reflex mapping without genuine learning. Practical tests on the isolated connectome show that the insect's architecture has strict hardware constraints, leading to catastrophic forgetting in complex tasks.
How the Digital Fly Brain Is Structured
Experiments rely on the onnxaa/fly-brain library, which contains a complete map of the Drosophila nervous system (~138,000 neurons and 15 million synapses). Most computational experiments focus on the mushroom body (MB)—a structure of roughly 6,300 neurons responsible for associative memory.
Unlike classical Reinforcement Learning (RL), where weights are gradually adjusted via a learning rate, the biological model operates via a binary dopamine trigger:
- Calling
step(odor=...)feeds an excitation vector into input ALPN neurons and returns an action/response score. - The
train(reward=..., punish=...)method applies a fixed-magnitude shift (~15%) to synaptic weights. Smooth scaling of reinforcement magnitude is not possible.
Multi-Armed Bandit and Number Comparison: Innate Preferences
In elementary multi-armed bandit tasks (selecting the most rewarding stimulus), the fly connectome performs on par with tabular Q-learning. However, the untrained biological network revealed strong innate asymmetry: by default, the model consistently preferred one stimulus over another. To level the playing field, forced preliminary exploration of each option (at least 8–10 steps) was required.
In magnitude comparison tasks, the connectome also demonstrated high accuracy (up to 95–100%), even successfully handling positional numeral notation and only beginning to struggle with 9-digit numbers.
The Maze: Channel Bottlenecks and Catastrophic Overfitting
Real problems arose when transitioning to sequential decision-making tasks. In mazes of various sizes, two critical barriers of biological wiring became evident:
| Maze size | Traversable cells | Connectome result (baseline) | Behavior with external gating |
|---|---|---|---|
| Small | 26 | Success (~99%, Q-learning level) | Not required |
| Medium | 72 | Success (~99%) | Not required |
| Large | 113 | Failure (~3.2% success, 295 steps per path) | 99% success from fixed start, drops to 75% with random start |
| Extra large | 316 | Complete collapse (0% success) | Brief spike up to 60%, rapidly falling to zero |
Reasons for degradation over larger distances:
- Hardware input limit: out of 685 input neurons, only 302 are activated per signal step. Encoding 452 cell-action pairs in a 113-cell maze required implementing sparse k-hot encoding.
- Memory erasure under continuous reinforcement: since the learning step size is fixed, repeatedly traversing familiar areas destroys weights for adjacent cells. This was partially mitigated by introducing a "gating" mechanism that suppresses
train()calls if the network already selects the correct action with a specified safety margin (0.05 margin). - Anatomical limit of separability: in a 316-cell maze, even abstract connectome representations performed worse (41–45%) than simple random noise (69–73%). Internal activation patterns for different cells overlap and can no longer be cleanly separated.
Tic-Tac-Toe and the Visual Pathway
Attempting to train the model to play tic-tac-toe confirmed the biological network's tendency toward self-destruction: in self-play mode, the baseline connectome's performance rapidly crashed to zero alongside a 19.6% weight drift. The situation was rescued only by artificially reducing the state space 6-fold through board symmetry.
The final test evaluated visual signal processing through visual receptors:
- Digit recognition (0–4): accuracy stalled at 45%. The connectome could not separate internal image codes (pattern overlap reached 77–84%) and collapsed most predictions into a single digit.
- Object counting: despite signals in the primary visual area correlating strongly with the number of shapes (r = 0.977), the mushroom body failed to transform this information into the correct action, achieving an accuracy of only around 20%.
The insect connectome is exceptionally well-suited for rapidly forming simple associations, but it is architecturally ill-equipped to retain hundreds of similar states or perform abstract processing of complex visual inputs without specialized evolutionary structures.