Can a Drosophila Whole-Connectome Network Learn Human-Designed Cognitive Tasks?

A wiring diagram tested outside its evolved repertoire

The question is whether a biological wiring diagram remains a useful computational substrate for problems the animal never faced. The authors take the publicly released MaleCNS v1.0 connectome, reconstructed from a single adult male Drosophila specimen, and freeze it as the recurrent topology of an artificial network. They then train two separate models on that same graph: one for bounded addition and one for a controlled grounded relational language task built from a fixed 100-word lexicon.

The design isolates topology from learning. In both models the only trainable tensor is a vector with one scalar per retained anatomical edge; input maps, readout maps, and token codebooks stay fixed. Synaptic contact counts determine only which edges exist, and edge scalars are initialized independently. So any advantage the anatomical graph shows cannot come from pretrained weights or from a richer parameter budget — it must come from how the edges are arranged.

The processed graph has 166,700 nodes and 25,582,938 directed neuron-to-neuron edges, representing 124,177,617 synaptic contacts. Because the comparison is against rewired graphs that keep each node's in-degree and out-degree, the study is really asking whether higher-order structure — reciprocity, motifs, paths, communities, spectra — adds anything beyond directed degree.

Rewires keep degree but erase almost every edge

The control graphs come from directed double-edge swaps that preserve every node's unweighted in-degree and out-degree, retain original self-loops, and introduce no duplicate edges. After five accepted swaps per original edge, only 0.588–0.594% of source edges remain across the 20-graph ensemble. Rewiring therefore changes the adjacency almost completely while holding node count, edge count, directed degree, task interface, parameter count, initialization distribution, optimizer, and training budget fixed.

That control is deliberately conservative, and the paper's own structural diagnostics show what it does and does not hold constant. Rewiring drops reciprocal directed edges from 29.893% to about 0.517% and shortens sensory-to-output minimum distance from 2.552 to 1.378, while the largest strongly connected component stays near 99.17% and every output node remains reachable. So the comparison tests the collective contribution of higher-order properties, not one isolated motif, and a simple reachability story is unlikely to explain the gap.

Addition accuracy separates the anatomical graph from rewires

The addition task contains every ordered pair of nonnegative integers with a + b ≤ 100, encoded in 47 fixed channels, with logits for the 101 answers averaged over the final four of 64 recurrent updates. The 5,151 expressions split into 3,584 training, 775 validation, and 792 test examples, and commuted expressions are grouped before splitting so no unordered operand pair appears in more than one split. Test accuracy therefore measures exact answers on unseen operand pairs.

Across completed runs, the anatomical graph averages 92.77% on the 792-expression test set, against 67.93% for the rewired graphs. The individual runs matter: anatomical runs range from 78.79% to 99.12%, while rewire runs range from 63.64% to 72.47%. The design is unbalanced — four runs on the anatomical graph, four on the first rewire, and two each on two additional rewires — so the authors report individual runs alongside a descriptive aggregate rather than a balanced factorial estimate.

The error measure points the same way. Mean absolute error is 0.124 answer-value units for the anatomical mean versus 0.404 for the rewire mean, and the anatomical graph scores 93.75% on the 20 expressions summing to exactly 100 versus 61.25% for rewires. Even the weakest anatomical run (78.79%) stays above every rewire run.

A strict paired endpoint rewards order-sensitive grounding

The language task pairs controlled English with task-defined sensory events. Each event combines one of six leg locations with one of four signal kinds and has binary strength, duration, and speed attributes; the 24 location–kind channels drive disjoint groups of anatomically annotated MaleCNS leg afferents comprising 2,837 neurons. The model must answer questions about held-out channel pairs and reversed event sequences, and orthogonal token codes carry no pretrained lexical semantics, so the mapping from sensory stream to words has to be acquired from the generated corpus.

The primary score is deliberately hard. For each held-out two-event record, the model must rank the correct description above the order-reversed description, and the reversed scene's correct description above the original one, with both cross-scene inequalities also satisfied; ties fail. With four exchangeable scores, only four of 24 orderings pass, giving an analytic chance reference of 1/6. The reported score is the proportion correct over 192 held-out groups.

On this endpoint the anatomical graph reaches 61.59% averaged across four fixed interfaces, against 44.17% for the matched rewire mean, a gap of 17.42 percentage points. The separation appears at every interface: 71.35% versus 44.92% at the canonical interface I1, 57.29% versus 41.02% at I2, 61.46% versus 46.09% at I3, and 56.25% versus 44.66% at I4. At I1 the anatomical graph also ranks first among a fixed 21-graph comparison, with a finite-ensemble rank fraction of 1/21; the 20 rewire scores there run from 36.98% to 58.85%, below the anatomical 71.35%.

Shuffling sensory features collapses the language score

A high score could in principle come from language priors rather than from reading the sensory scene. The sensory intervention tests that directly. On a matched 48-group subset, the full sensory-amplitude tensor is replaced by that of the next group in a deterministic cycle, while the prompt and candidate answers stay unchanged, so both scores are computed on the same records.

Under that shuffle, the anatomical model's score falls from 60.94% to 19.27%, and the rewire mean falls from 41.80% to 13.02%. The drop shows the score depends on the specific sensory input rather than on answer statistics alone. The paper also reports that both no-sensory checkpoints score 0% on the primary and counterfactual endpoints and under sensory shuffle.

Grounding is uneven across attributes

The attribute-level counterfactuals show that the sensory-to-text mapping is not uniform. Using pair-preserving substitutions on 48 groups per factor, the anatomical model reaches 46.35% two-preference accuracy for strength changes and 59.38% for duration changes, but only 4.17% for speed changes; order changes reach 26.56%. The macro mean is 34.11% for the anatomical model versus 8.66% for rewired graphs.

Under the stricter four-inequality criterion the same pattern holds at lower levels: 17.19% for strength, 22.40% for duration, 0.00% for speed, and 26.04% for order, with a macro mean of 16.41% versus 4.69% for rewires. The authors read this as evidence that the anatomical model distinguishes strength and duration more reliably than speed. Location and signal-kind substitutions are excluded from this analysis because they may move a record into a different held-out channel-pair family.

The reusable asset is architecture, not trained weights

Because addition and language are fitted independently, nothing learned in one task is transferred to the other. What carries across is the recurrent architecture itself. The authors describe this as higher-order MaleCNS wiring providing a reusable inductive bias for both bounded addition and grounded relational language, and they note that the language gap persists across token-input and output placements and across graph draws while sensory afferents remain in their anatomical groups.

The paper also reports a dense control whose results sit far below the connectome model: across four optimization runs, strict paired accuracy averages 8.72%, normal 9.38%, shuffled 1.56%, and counterfactual 3.39%. That control matches recurrent parameter count rather than state dimension or sparsity, a distinction the authors flag themselves, so it should not be read as a like-for-like architectural comparison.

One graph, two synthetic tasks, and a compact vocabulary

The evidence rests on a single anatomical graph and two synthetic tasks. The addition comparison is an unbalanced graph-by-initialization design, and the 20-graph language ensemble limits rank resolution to 1/21. Four interfaces sample only a small part of the possible placement space. The compact vocabulary, authored semantics, and controlled grammar make relational generalization measurable but leave open how the result scales to natural language.

The network also omits synaptic signs, neurotransmitter identities, delays, cell dynamics, and biological plasticity, so it is a topological abstraction rather than a model of the fly nervous system. Within those bounds, the paper's most decision-relevant comparison is not the headline accuracy but the gap against degree-preserving rewires: the anatomical graph wins on both tasks, at every tested language interface, and its language score depends on the actual sensory input rather than on answer statistics. The practical priority the evidence supports is testing whether null models that preserve reciprocity, community structure, or sensory-to-output path statistics can account for that gap, since those are the properties rewiring most strongly altered.

Sources used in this answer