Hemibrain v1.2 130 projection neurons 1,745 Kenyon cells 9,657 synaptic connections --

The Fly's Hash Function

In 2017 the fruit fly's olfactory circuit was found to compute a locality-sensitive hash. The wiring looked random, so every paper since has modelled it with a random matrix. We now have the real wiring. Below, the measured connectome runs live against that assumption — same input, same inhibition, only the synapses differ.

Antennal lobe → mushroom body calyx Signal: drifting blend
Projection neuron input — 130 channels, 52 glomeruli  
Real connectome measured synapse counts
Active— Similarity kept —
Random wiring the 2017 assumption
Active— Similarity kept —
Odour
Input world Statistics matched to each wiring's own structure
Both circuits are hashing the same signal. Similarity kept is the fraction of the sparse code that survives a small perturbation of the input — the property a locality-sensitive hash exists to provide.

What the measured wiring actually does

Each Kenyon cell samples about six projection neurons through dendritic claws, and feedback inhibition leaves only the strongest 5 % firing. Those survivors are the hash code. Swapping the real matrix for matched null models isolates what each structural feature costs.

Retrieval quality — precision@16, five seeds

Fraction of each point's true 16 nearest neighbours recovered from the sparse code alone. Every condition shares one fixed encoder into projection-neuron space, so only the 130×1745 matrix varies.
The real connectome is ~20 % worse at generic similarity search than the random matrix standing in for it. Nine years of fly-inspired hashing has been crediting the fly with performance its own wiring does not have.

Where the cost comes from

The nulls separate cleanly. Uneven projection-neuron out-degree is the dominant penalty; the structured co-occurrence reported in 2022 costs almost nothing; synaptic weighting costs nearly as much as degree. Strip all three and you recover the idealised circuit exactly.

But it is specialised, not simply worse

Feeding the circuit input drawn from the world its own wiring implies reverses the result completely. This is the switch wired into the console above — flip between the two worlds and watch the winner change.

Rows are the input distribution, columns the matrix doing the hashing. Each world is sampled with covariance equal to that wiring's projection-neuron co-occurrence.

So the mushroom body is not a poor hash function. It is a hash function tuned to one distribution, paying for that tuning everywhere else — and the tuning lives almost entirely in the degree sequence rather than in which specific glomeruli share a cell.