Source-linked AI summary
A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks
Joonghui Cho, Minchan Kang, Daeshik Kim
TL;DR
This paper asks whether connectome organization remains useful for computational tasks beyond an animal’s natural repertoire, especially given mixed evidence about whether biological topology itself matters. It trains separate recurrent models on the MaleCNS topology for bounded addition and grounded relational language, and finds that the anatomical graph outperforms degree-preserving rewires on both tasks. These results support a reusable architectural role for higher-order MaleCNS wiring within the tested settings.
Problem
Whether biological edge organization remains useful for tasks beyond an animal’s natural repertoire is an open question, and prior topology comparisons give mixed evidence.
Source: §1 · p. 1Method
Results
The anatomical graph reaches 92.77% mean held-out addition accuracy versus 67.93% for directed degree-preserving rewires, and 61.59% strict paired language accuracy versus 44.17% for matched rewires.
Source: §4.1 · p. 4, §4.2 · p. 4Takeaways & Limitations
Across the tested tasks, interfaces, and graph draws, the results identify the connectome as a useful architectural prior whose recurrent architecture transfers across independently fitted tasks.
Source: §5 · p. 5, §7 · p. 6The study evaluates one anatomical graph and uses a compact vocabulary, authored semantics, and controlled grammar, leaving scaling to natural language open.
Source: §6 · p. 6
Abstract
from arXiv · showhide
Can a biological wiring diagram serve as a useful computational substrate beyond the behaviors for which it evolved? We use the publicly released MaleCNS v1.0 connectome, reconstructed from a single adult male Drosophila specimen, as the fixed recurrent topology of an artificial network. We train separate models for bounded addition and for a controlled grounded relational language task built from a fixed 100-word lexicon. In both models, one scalar is learned per anatomical edge. The anatomical graph reaches 92.77% mean accuracy on held-out addition, compared with 67.93% for directed degree-preserving rewires. On the strict paired language endpoint, which matches original and order-reversed scenes to their corresponding descriptions, it reaches 61.59% across four fixed interfaces, compared with 44.17% for matched rewires. At the canonical interface, it ranks first in a fixed 21-graph comparison. On the matched 48-group intervention subset, shuffling task-defined sensory features reduces its score from 60.94% to 19.27%. Together, these results show that higher-order MaleCNS wiring provides a reusable inductive bias for bounded addition and grounded relational language.
1 Introduction
Connectomes may provide candidate architectures for tasks beyond an animal’s natural repertoire, but task performance alone cannot establish whether biological edge organization matters. This study tests a MaleCNS-based recurrent network on separately trained addition and grounded relational language tasks.
1 Introduction
- The study asks whether MaleCNS organization remains useful when connection strengths are relearned for problems outside the animal’s natural repertoire.Connectome reconstructions cover nearly the entire adult fly brain and the complete adult male CNS.Source: §1 · p. 1
- Prior topology comparisons found similar performance for random and biological structures in one study, while degree-preserving nulls removed much of an apparent advantage in another.These mixed findings mean task performance alone does not show whether biological edge arrangement matters.Source: §1 · p. 1
- The models use the MaleCNS adjacency as a recurrent mask, with one learned scalar per retained edge, and are fitted separately for addition and grounded relational language.Addition tests exact answers on unseen operand pairs; language tests held-out channel pairs and reversed event sequences using controlled English and task-defined sensory events.Source: §1 · p. 1
- The anatomical graph outperforms degree-preserving rewires on both tasks, with the language separation appearing at every tested interface and in an additional 20-rewire ensemble.Figure 1 outlines the fixed recurrent support, separate task fits, and main controls.Source: §1 · p. 1
2 Related Work
Prior work has used connectomes for biological modeling and artificial tasks, while studies of topology’s computational value report mixed findings. This paper positions its experiments as symbolic tests of connectome architecture across tasks, interfaces, and graph draws.
Whole-connectome resources and models
- Whole-connectome studies have modeled visual activity, feeding and grooming transformations, causal neural effects, and resting-state activity and connectivity.Fly-Wire reconstructed a female brain with 139,255 neurons and about 50 million chemical synapses; MaleCNS extends synapse-resolution reconstruction across the complete adult male CNS.Source: §2 · p. 1
Topology as an inductive bias
- Topology studies report higher weight-cost-normalized performance and robustness in small fly subgraphs, sensitivity to degree-preserving rewiring in communication models, and advantages that can disappear with stronger null controls.The cited fly-visual-system study found advantages over a weak random control largely disappeared with shared from-scratch initialization and degree-preserving nulls.Source: §2 · p. 2
Artificial-task studies
- Artificial-task studies have used connectomes for image classification, chess, simulated fly locomotion, and robot navigation, including a locomotion comparison with a degree-preserving rewire.The present experiments instead test two symbolic mappings across token-input and output placements and graph draws.Source: §2 · p. 2
Grounding, composition, and arithmetic
- Related language work spans sparse word representations, learning word-referent associations, pretrained instruction embeddings, and benchmarks of compositional generalization.The paper’s language experiment uses orthogonal token codes, channel-pair holdouts, sensory interventions, and relation reversal; its addition experiment reports exact answers and carry-sensitive results.Source: §2 · p. 2
3 Methods
The methods fix the MaleCNS recurrent topology and compare it with directed degree-preserving rewires while training separate addition and grounded-language models. Evaluation combines held-out task protocols with interface, sensory-shuffle, and graph comparisons.
3.1 Experimental design
- Five topologies and four language interfaces are crossed, with a further 21-graph comparison at canonical interface I1; rewiring preserves node and edge counts, directed degrees, and training conditions.The language and addition tasks are fitted independently.Source: §3.1 · pp. 2–3
3.2 Connectome network and topology controls
- The processed graph contains 166,700 nodes and 25,582,938 directed edges; each retained edge has a learned scalar, while rewires preserve each node’s in- and out-degree.Double-edge swaps retain self-loops and avoid duplicate edges; after five accepted swaps per original edge, only 0.588–0.594% of source edges remain across the 20-graph ensemble.Source: §3.2 · p. 3
3.3 Bounded-addition protocol
- Addition covers ordered nonnegative operand pairs whose sum is at most 100, with commuted pairs grouped before splitting into training, validation, and test sets.Each epoch visits all training expressions once; the checkpoint with lowest validation loss is evaluated once on the test set, across an unbalanced design of four anatomical runs and eight rewire fits.Source: §3.3 · p. 3
3.4 Grounded relational language protocol
- The language task pairs sensory events with controlled English and fixed orthogonal token codes, using disjoint sensory, token-input, and output node sets.Training retains the sensory sequence’s recurrent state while reading the prompt; held-out evaluation requires both original and order-reversed scenes to match their corresponding descriptions, and the primary endpoint counts a group correct only if all four score comparisons are strict.Source: §3.4 · pp. 3–4
3.5 Controls and statistical analysis
- Language controls include a 48-group sensory-shuffle intervention and a parameter-matched dense recurrent model, while the factorial analysis estimates within-interface differences against four rewires.A separate I1 analysis ranks the anatomical graph against 20 rewires; two additional initializations test optimization robustness.Source: §3.5 · p. 4
4 Results
The anatomical graph outperforms degree-preserving rewires on held-out addition and grounded relational language. Language performance also declines sharply when task-defined sensory amplitudes are shuffled.
4.2 The topology advantage extends to grounded relational language
- 61.59% on the strict paired language endpoint across four interfaces compares with 44.17% for rewires, a 17.42 pp difference.The 95% crossed-bootstrap interval is [11.39, 24.90] pp, and the separation is positive at every tested placement.Source: §4.2 · p. 4, Table 1 · p. 5
4.1 The connectome topology improves held-out addition
- 92.77% mean held-out addition accuracy on the anatomical graph compares with 67.93% on rewires, a 24.84 pp difference.For examples with a ones-digit carry and sum below 100, the means are 92.96% and 69.09%, respectively.Source: §4.1 · p. 4, Figure 2 · p. 5
4.2 The topology advantage extends to grounded relational language
- At I1, the anatomical graph scores 71.35%, while none of 20 independent rewires reaches that score.The rewires average 46.56% (SD 6.99 pp; range 36.98–58.85%).Source: §4.2 · p. 5
- At I1, three anatomical language runs average 71.01% (SD 2.62 pp), compared with 46.53% for Rewire 1 and 46.18% for Rewire 2.Language optimization variance is smaller for the anatomical runs than for either rewire, while anatomical addition runs show a wider spread.Source: §4.2 · p. 5
- Both topology classes have 99.17% of nodes in their largest strongly connected component and a mean token-to-output distance of 1.51, but reciprocity is 29.89% for the anatomical graph versus 0.514–0.520% for rewires.Mean minimum sensory-to-output paths are 2.55 for the anatomical graph and 1.38 for rewires.Source: §4.2 · p. 5
4.3 Sensory alignment and relational generalization
- On the 48-group intervention subset, shuffling sensory amplitudes reduces the anatomical score from 60.94% to 19.27%, a 41.67 pp drop.The mean rewired score falls from 41.80% to 13.02%, a 28.78 pp drop.Source: §4.3 · p. 5
- On the two-preference counterfactual endpoint, the anatomical graph reaches 34.11% versus 8.66% for rewires; strict four-comparison means are 16.41% and 4.69%.The parameter-matched dense recurrent model scores 3.39% on the two-preference endpoint.Source: §4.3 · p. 5
5 Discussion
The anatomical graph’s advantage indicates that directed degree alone does not capture all useful connectome organization, while language performance also depends on sensory-text correspondence. Structural diagnostics make simple reachability an unlikely explanation, but leave the contributions of other network properties unresolved.
5 Discussion
- Both topology classes retain input-to-output access, making simple reachability an unlikely explanation for the performance gap.Null models preserving reciprocity, community structure, or sensory-to-output path statistics could distinguish their contributions.Source: §5 · p. 5
- On the same 48 groups, sensory shuffling reduces the anatomical score from 60.94% to 19.27% and the rewire mean from 41.80% to 13.02%.The anatomical model distinguishes strength and duration changes more reliably than speed changes.Source: §5 · p. 5, Figure 3 · p. 6
- At I1, the anatomical graph scores 71.35%, above all 20 degree-preserving rewires and their 46.56% mean.Its finite-ensemble upper-tail rank fraction is 1/21.Source: §5 · p. 5, Figure 4 · p. 6
6 Limitations
The evaluation is limited to one anatomical graph and two synthetic tasks, and its language tests sample only a small part of the possible interface placements. The controlled language setup and simplified network also leave natural-language scaling and several biological features untested.
6 Limitations
- The study evaluates one anatomical graph and two synthetic tasks.The addition runs use an unbalanced graph-by-initialization design, while the 20-graph language ensemble limits rank resolution to 1/21.Source: §6 · p. 6
- Four interfaces sample only a small part of the possible placement space.Source: §6 · p. 6
- The compact vocabulary, authored semantics, and controlled grammar leave scaling to natural language open.The network omits synaptic signs, neurotransmitter identities, delays, cell dynamics, and biological plasticity; its dense control matches recurrent parameter count rather than state dimension or sparsity.Source: §6 · p. 6
7 Conclusion
Separate fits on processed MaleCNS topology learn bounded addition and controlled grounded relational language. Degree-preserving rewiring reduces performance on both tasks, and sensory shuffling sharply reduces language performance, supporting the connectome as a useful architectural prior.
7 Conclusion
- Separate fits on processed MaleCNS topology learn bounded addition and controlled grounded relational language.Source: §7 · p. 6
- Degree-preserving rewiring reduces performance on both tasks, with the language gap persisting across tested token-input and output placements and graph draws.Source: §7 · p. 6
- Sensory shuffling sharply reduces language performance, linking learned descriptions to the task-defined input stream.Source: §7 · p. 6
- The results identify the connectome as a useful architectural prior and motivate isolating the structural properties behind its advantage.Source: §7 · p. 6
AI Use Statement
Generative AI tools were used to develop and review the synthetic data generator and polish manuscript prose; their use in other listed research components was limited to review. The authors take responsibility for the final content.
AI Use Statement
- Generative AI tools helped develop and review the synthetic data generator and polish the manuscript’s prose.Source: AI Use Statement · p. 6
- For training code, equations, literature, methodology, result interpretation, figures, and translations, AI use was limited to review.Source: AI Use Statement · p. 6
- The authors take responsibility for the final content.Source: AI Use Statement · p. 6
A Data and Graph Provenance
The paper constructs a directed graph from MaleCNS v1.0 and uses fixed input, output, and graph assignments with learned edge scalars. The graph comes from one adult male specimen.
A Data and Graph Provenance
- The processed graph retains 166,700 body IDs; the source connectivity graph has 166,483 neurons after excluding 217 without synaptic connections.Repeated records for an ordered neuron pair are combined by summing contact counts; counts determine edge presence, not initial weight magnitudes.Source: A Data and Graph Provenance · p. 8, Table 2 · p. 8
- Rows are postsynaptic and columns presynaptic, so edge (i, j) carries input from neuron j to neuron i.The 101 self-edges remain unchanged under rewiring.Source: A Data and Graph Provenance · p. 8
B.1 Shared recurrent parameterization
- Only edge scalars are optimized; the input groups, output groups, graph indices, and language codebooks remain fixed.Edge scalars are independently initialized from N(0, 1) using a run-specific random state, and recurrent states begin at zero.Source: §B.1 · p. 8
B.2 Addition task and optimization
This section specifies the bounded-addition and controlled language tasks, their graph interfaces and evaluation procedures, and the rewired and dense comparison models. It also reports structural and counterfactual findings, with detailed run results assigned to tables whose cell values are not included here.
B.2 Addition task and optimization
- Addition uses grouped input and answer nodes, with temperature 0.1 and a loss combining cross-entropy with a normalized ordinal CDF penalty.The penalty term targets disagreement between predicted and target cumulative distributions over ordered answer classes.Source: §B.2 · pp. 8–9
B.3 Language task specification; B.4 Corpus generation and split construction
- The language generator combines event reports, attribute and relation questions, and controlled text; composition tests hold out channel pairs and include reversed event sequences.The corpus contains 200,000 training records and 119,040 distinct conditional examples after deduplication.Source: §B.3 · p. 9, §B.4 · p. 10
B.5 Language interface and optimization
- Language inputs combine 24 groups of anatomical afferents with token codes; the network encodes sensory events, then reads the prompt and predicts an answer.Primary results use the final checkpoint after a common four-million-token budget.Source: §B.5 · p. 10
C.1 Strict paired language endpoint
- A strict paired group counts as correct only when both matched scene-description scores exceed both mismatched scores; ties fail.The primary endpoint evaluates 192 held-out groups, and its two-preference counterfactual counterpart tests changes to strength, duration, speed, or order.Source: §C.1 · p. 11
B.6 Hyperparameter selection and interfaces; C.2 Factorial uncertainty and graph ranking
- The language comparison averages anatomical-versus-rewire differences across four interfaces and uses a crossed bootstrap to estimate variation within that executed design.A separate I1 analysis ranks the anatomical graph against 20 rewires without assuming uniformity of the double-edge-swap chain.Source: §C.2 · p. 11
D Rewire Construction and Structural Diagnostics
- Rewiring preserves the giant strongly connected core and token-to-output distance nearly unchanged, but greatly reduces reciprocity and shortens sensory-to-output paths.The structural differences motivate testing null models beyond shared reachability.Source: D Rewire Construction and Structural… · p. 11, Table 7 · p. 12
D Rewire Construction and Structural Diagnostics; E Compute and Reproducibility
- MaleCNS v1.0 source tables and skeletons are available from the official release page, while the project repository provides code, compact result records, checksums, and instructions for rebuilding the graph and generated corpus.MaleCNS data and full training checkpoints are not redistributed.Source: E Compute and Reproducibility · p. 12
F.2 Language factorial and controls; B.7 Language controls
- Counterfactual accuracy is concentrated in strength and duration, while speed remains near zero; both no-sensory checkpoints score 0% on the primary and counterfactual endpoints and under shuffle.The no-sensory controls disable sensory input during both training and evaluation.Source: §B.7 · p. 11, §F.2 · pp. 13–14