A Connectome Test of the Fly Hashing Algorithm

Sebastian Senge

cs.NEOct 6, 2026

A Connectome Test of the Fly Hashing Algorithm

Authors: Sebastian Senge

Organizations: Independent Researcher

Abstract

Dasgupta, Stevens and Navlakha (2017) showed that the Drosophila olfactory circuit, modelled as a random sparse projection followed by winner-take-all, is a locality-sensitive hash that beats classical LSH. The projection was random because the wiring was unknown. We test it against four electron-microscopy connectomes (MaleCNS, hemibrain, FlyWire, BANC; four animals, seven hemispheres). First, the 2017 pattern holds in a reimplementation of its protocol on SIFT, MNIST and odour mixtures (GloVe is near chance at short codes for every method): the fly hash beats k Gaussian projections at short hash lengths (3.1x in AP@200 on MNIST at k = 4). Second, against the tested real-valued Gaussian baseline that advantage is per active cell, not per operation: Gaussian projections given the same projection arithmetic retrieve better on every dataset and input dimension tested. Third, across four connectomes the measured pairing of glomeruli gives no consistent retrieval advantage over degree-preserving rewiring: retrieval is slightly lower (median -1.6%), and the small odour deficits depend on how missing odour responses are treated. Separately, equalising glomerular fan-out at fixed connection count improves retrieval in the model in every hemisphere, while equalising inputs per cell lowers it on average. Yet the fan-out profile is similar across the four sampled animals (median between-animal Spearman rho = 0.86), structural synapse counts do not offset it, and its relation to odour tuning is weak. In the model that uneven allocation costs retrieval. For practice: measured wiring gives no consistent retrieval advantage over degree-preserving random wiring, so a fly hash needs no connectome data, and its advantage is per active unit, which may suit hardware where active units rather than arithmetic are the binding cost, a hypothesis we do not test.

Figures & tables

Fig. 1: (a) The fly hash: glomerular input

x

is expanded through a binary projection

M

onto Kenyon cells and the

k

most active cells form the tag. (b) Glomeruli reach very different numbers of Kenyon cells, and the profile is similar across the sampled hemispheres: relative fan-out of the

49

shared glomeruli in all seven, ordered by the mean; the dashed line is even fan-out. (c) Inputs per Kenyon cell, against the six of the 2017 construction.

element	choice	source
data	$10{,}000$ vectors: SIFT-small base, GloVe 6B 300d and MNIST train (seeded subsets)	2017 (size); ours (subset)
queries, truth	$1{,}000$ queries; top $2\%$ by Euclidean distance on raw features	2017
preprocessing	each vector centred by its own mean	2017
fly	$m=20k$ or $10d$ ; top- $k$ winners	2017
sampling	each cell samples $10\%$ of inputs	2018 code
tag	binary; Hamming distance, random tie-break	2018 code; ours

TABLE I: Provenance of the protocol: 2017 main text [ 3 ] , 2018 FlyLSH code [ 20 ] , or our choice.

Fig. 2: The 2017 protocol with random matrices: AP@200 against hash length, mean

\pm

SD of 50 trials. Random tag selection performs like LSH; winner-take-all at

m=20k

and

10d

beats both at short hash lengths.

$k$	fly	sign, $k$ bits	storage-matched	operation-matched
8	0.197	0.019	0.264 (72)	0.305 (102)
16	0.288	0.069	0.340 (130)	0.305 (102)
32	0.353	0.151	0.394 (229)	0.305 (102)
64	0.396	0.249	0.436 (397)	0.305 (102)
92	0.406	0.296	0.454 (523)	0.305 (102)

TABLE II: Odour analysis, MaleCNS: AP@10 of the fly hash and of one-bit Gaussian codes (bits in parentheses).

Fig. 3: (a) One dataset at several input dimensions (MNIST reduced by PCA to

d

; PCA also changes the retained information and the integer sampling rate),

k=4

: the fly hash’s lead over LSH per active cell grows with

d

, but LSH given the same projection arithmetic as the

m=10d

fly leads at every

d

. (b, c) Best AP of the fly hash over the tested hash lengths against real-valued LSH with the same projection arithmetic: (b) random matrices on the original inputs, (c) the MaleCNS connectome on its own inputs. Inputs and ground truth differ between (b) and (c); only pairs within a panel are comparable.

$k$	conn.	null	$\Delta$ %	$p$	$p_{\text{Holm}}$	fan flat	6 unif.	balanced
8	0.1973	0.2034 $\pm$ 0.0085	-3.0	0.468	1.000	+3.7	+26.7	+3.2
16	0.2879	0.2933 $\pm$ 0.0062	-1.8	0.438	1.000	+2.7	+9.7	+3.2
32	0.3531	0.3575 $\pm$ 0.0045	-1.2	0.338	1.000	+3.3	+3.0	+3.6
64	0.3963	0.4001 $\pm$ 0.0039	-0.9	0.348	1.000	+3.4	+0.7	+3.4
92 ∗	0.4063	0.4148 $\pm$ 0.0035	-2.1	0.020	–	+3.2	+0.1	+3.2

TABLE III: Odour analysis, MaleCNS: connectome against the null (

B=200

) and comparison matrices (relative to the null mean, %). ∗ primary.

variation	$\Delta$ %	$p$
Baseline	-2.2	0.049
Left hemisphere	-1.1	0.244
$\geq$ 5-synapse threshold	-2.3	0.098
Missing data (3 imputations)	-0.7 to +0.6	0.439–0.780
Coverage thresholds (2)	-1.7 to -1.4	0.049–0.341
All 46 glomeruli with data, imputed	-0.0	1.000

TABLE IV: Odour analysis, MaleCNS, primary size, one-factor variations (

B=40

nulls each): relative difference (%) and two-sided

p

Fig. 4: Connectome against its own null (AP relative to the null mean). (a) Odour analysis, MaleCNS right: orange, 200 null matrices; blue, the connectome. (b) The 2017 protocol in seven hemispheres: four datasets

\times

six hash lengths each, bar at the median. (c) Odour analysis at the primary size in every hemisphere, on its own DoOR glomeruli and on the

34

all share.

hemisphere	sex	cells	inputs	fan CV	$Q$ $z$	odours $\Delta$ % ( $p$ )	shared $\Delta$ % ( $p$ )	protocol median $\Delta$ %	fly/LSH
MaleCNS R	m	1,886	5.28	0.53	+2.7	-2.2 (0.015)	-2.6 (0.020)	-1.3	2.3
MaleCNS L	m	1,865	5.30	0.54	+4.8	-0.9 (0.294)	-0.8 (0.376)	-1.0	2.2
Hemibrain R	f	1,761	6.05	0.50	+0.0	-0.7 (0.491)	-0.0 (0.990)	-1.1	2.3
FlyWire R	f	2,406	5.00	0.47	+11.3	-2.7 (0.001)	-2.7 (0.020)	-0.7	2.4
FlyWire L	f	2,378	5.12	0.49	+17.8	-1.6 (0.037)	-1.6 (0.079)	-2.0	2.9
BANC R	f	2,046	5.03	0.57	+6.2	-3.1 (0.003)	-3.0 (0.020)	-3.2	2.3

TABLE V: Seven hemispheres of four animals: Kenyon cells, mean inputs per cell, fan-out coefficient of variation, co-occurrence

Q

z

against

1000

nulls; odour analysis at the primary size on own glomeruli (

2000

nulls, floor

0.001

) and on shared glomeruli (

100

nulls) (relative difference %, unadjusted

p

); median over the 2017-protocol runs (

4

datasets

\times

6

sizes,

B=50

, 5 trials); fly hash over LSH on MNIST at

k=4

Fig. 5: Equal-connection degree controls, AP@200 relative to the null mean (ratio of trial means), with

95\%

bootstrap intervals over trials, each trial paired with its own null mean. (a) MaleCNS right at the largest hash size (20 trials,

10

nulls). (b) All hemispheres, averaged over the four datasets and six hash sizes.

task	null	conn.	fan-out even	inputs even	both	six
Novelty, MNIST (AUC)	0.741	-0.006	+0.003	-0.015 †	-0.012 †	-0.009
FlyNN-style, MNIST (accuracy)	0.700	-0.009 †	+0.010 †	-0.009 †	-0.001	+0.013 †
Novelty, odours (AUC)	0.560	-0.003	-0.007 †	-0.001	-0.007 †	-0.002

TABLE VI: Novelty detection and nearest-neighbour classification (MaleCNS right, 10 trials): null mean, and difference to it for the connectome, the equal-connection controls and the six-input construction. † 95% interval excludes zero.

Explore similar work

Null-model treatment of the sensory-motor boundary changes an evolutionary connectome comparison
Sep 30, 2026Gyujeong ParkHuman Connectome Project Connectome
IonLabs
Structure alone supports efficient visual computation in the Drosophila visual system
Oct 7, 2026Eudald Correig-Fraga, Roger Guimerà, Marta Sales-PardoVisual Perception
Innovamat Education, Sant Cugat del Vallès, Catalonia · Department of Chemical Engineering, Universitat Rovira i Virgili, Tarragona, Catalonia · Center for Computational Science and Applied Mathematics (ComSCIAM), Universitat Rovira i Virgili, Tarragona, Catalonia +1
A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks
Oct 7, 2026Joonghui Cho, Minchan Kang, Daeshik KimCognitive Science
Korea Advanced Institute of Science and Technology (KAIST)

Sep 30, 2026, cs.NEJ/K move · Enter open · S save

Null-model treatment of the sensory-motor boundary changes an evolutionary connectome comparison

Gyujeong Park

IonLabs

Human Connectome Project Connectome Sensorimotor Experience Smells Random Heterogeneous Treatment Effects+1

Open paper AskPDF arXiv

Randomised copies of a connectome are the usual baseline for asking whether measured wiring matters, and the answer depends on what the randomisation preserves. We evolved embodied foraging agents whose brains are a compressed adult Drosophila connectome (FlyWire v783; 512 cell-type groups and 1,000 Kenyon cells) alongside agents built on randomised wiring, in pre-registered experiments with ten seeds, four ecologies and 600 generations. Two standard randomisations, a column shuffle and degree-preserving edge swaps, route 10.6 to 10.7 % of olfactory output directly onto descending motor groups, against 0.012 % in the connectome. On the registered primary endpoint, fitness averaged over the run, no difference was detected; at the last common-garden probe the connectome was behind both controls (-0.22 and -0.20 fitness units on seed means). Against controls that keep every sensory-output and motor-input edge and rewire only the interior, the seed-mean difference lay within a +/-0.10 equivalence bound (+0.002 and -0.074, unchanged under a calibration that also matches activity spread), although per ecology the interior column shuffle was ahead by 0.26 in one of four ecologies at ten seeds, a lead that ten further pre-registered seeds did not replicate. Rewiring the connectome so that it acquires the shortcut raised its fitness by 0.44 (10 of 10 seeds) and its dependence on olfaction from 0.15 to 0.99; graded doses raised both in step; at comparable swap counts the full dose was ahead of an interior-only sham by 0.53 (10 of 10 seeds); and a sham that rewired the same boundary edges without creating shortcuts matched the connectome (+0.007) while the full dose was ahead of it by 0.60. What a null preserves at the sensory-motor boundary can decide an evolutionary connectome comparison, and sensory-to-motor path statistics belong next to the degree statistics a null is said to preserve.

Abstract

Explore similar work

Null-model treatment of the sensory-motor boundary changes an evolutionary connectome comparison

Structure alone supports efficient visual computation in the Drosophila visual system

A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks