Berg et al. (Janelia FlyEM, Google Research, and collaborators) reconstructed the entire central nervous system of an adult male Drosophila — brain, both optic lobes, and the ventral nerve cord — as one connected graph: 166,691 neurons, ~25.6 million directed connections, 46M presynapses / 312M postsynaptic densities, 11,691 cell types. It is the largest brain map by neuron count to date, the first to include the nerve cord (not just the brain), and the first male connectome, enabling a synaptic-resolution male–female comparison. Read as an artificial neural network, it is an instructive counter-example to almost every default of deep learning.
The graph, not the stack
The fly CNS is not a layered, densely connected network. It is a sparse, directed graph: nodes are neurons, edges are ordered neuron pairs (A→B differs from B→A in both existence and strength). A fully connected graph over 166,700 nodes would need ~27.8 billion edges; the brain uses ~25.6 million — a density of ~0.02%, with each neuron talking to ~37 partners on average. Yet it is not fragmented: >90% of neurons lie in one giant reachable component with short path lengths, and nearly every neuron sits inside a recurrent loop. This is the topology of a graph neural network, and the ~65% recurrence measured in the larva (winding-2023-the-connectome-of-an-insect) recurs here at 50× scale.
Weights as fractional share, and coincidence as computation
Synapse count sets connection strength, but functionally as a share of the target’s total input, not an absolute weight. If 15 of neuron B’s 100 incoming synapses come from A, then A holds a 15% “vote” in B. Crucially, no single input crosses threshold alone — the neuron fires only on coincident input from several partners. The circuit reads simultaneity, not individual signals. Inhibitory inputs enter the same weighted sum as negative terms, so excitation and inhibition compete in one addition — the McCulloch–Pitts / perceptron thresholded sum (mcculloch-1943-a-logical-calculus-of-the, rosenblatt-1958-the-perceptron-a-probabilistic-model), but embedded in a sparse recurrent graph rather than dense layers.
Architecture pre-trained by evolution
An artificial network has a human-designed architecture and randomly initialized weights. The fly inherits the opposite: who connects to whom, and with how many synapses, is largely genetically specified — a prior found by evolution’s slow outer loop, the same principle already visible in C. elegans (white-1986-the-structure-of-the-nervous) and the larval brain. Large parts of sensory processing (vision, olfaction) behave like frozen weights: the ability to recognize an odor is built in at birth; what a fly learns in minutes is the thin layer downstream that decides whether to approach or avoid. This is transfer learning’s frozen backbone plus a small trainable head — and it is why perception does not change when behavior does (the burned hand still sees the iron perfectly; it just won’t reach for it).
What the map does not contain
A connectome records wiring, not learned efficacy. Biological weight ≈ synapse count × per-synapse efficacy, and only the count is in the map; learning largely adjusts the efficacy (release probability, receptor sensitivity) while leaving the wiring intact. So two flies that learned opposite associations would produce near-identical connectomes. What we have obtained is not what a fly learned, but where learning can happen — the structure, not the state.
Significance for AI
- Small, sparse, sufficient: 166,700 neurons and 5-bit-scale synaptic resolution suffice for walking, flying, courtship, and one-shot learning. The lesson the fly offers is not how to build bigger models, but how a small model learns and acts with little computation.
- Structure matters, independent of weights: connectome-constrained network simulations of the earlier female brain learned to walk, fly, and solve tasks; a control that shuffled the edges while keeping the weights learned more slowly — evidence that the inherited graph itself carries computational value.
- A biological baseline for GNNs and quantized, recurrent, mostly-frozen networks: the fly CNS is the most complete existing example of the regime modern efficient-ML keeps rediscovering — sparse graphs, heavy recurrence, frozen backbones with tiny trainable heads, and coarse weight resolution.
Related connectomes
| Organism | Neurons | Synapses | Status |
|---|---|---|---|
| C. elegans | 302 | ~7,000 | Complete (1986) |
| Drosophila larva | 3,016 | 548,000 | Complete (2023) |
| Drosophila adult female brain | ~140,000 | ~50M | Complete (2024) |
| Drosophila adult male CNS (brain + VNC) | 166,691 | ~358M sites | Complete (2026) |
| Mouse | ~71M | ~100B | Partial |
| Human | ~86B | ~100–500T | Not started |