acceptodds
Under review as a conference paper at ICLR 2027

Learning from the Observation Graph of Incomplete Connectomes

Abstract

Connectome reconstruction records synaptic relations before every observed segment has a resolved neuronal identity. Restricting inference to traced neurons can therefore discard evidence that is already available. We formulate learning on this evolving observation graph and distinguish the value of preserving unresolved relations from the additional value of learning to combine them. Our leave-fragment-out supervision samples count-conserving queries from traced source neurons and subtracts each query from its positive owner before feature construction, removing direct query self-overlap. Models trained in MaleCNS are frozen and evaluated on recorded MANC revisions: 557 donor segments in 35 components, each ranked against all 23,200 earlier traced candidates. The segment-aware neural scorer achieves component-macro Hit@20 of 0.5399, compared with 0.4866 for its retained-only counterpart and 0.4666 for fixed cosine. Gains depend on the inspection budget: Hit@1, Hit@5, and mean reciprocal rank decrease. A separate masked-adjacency task improves from 0.3800 to 0.4708 average precision when unresolved context is retained. Correspondence interventions and support-stratified analyses show where that context contributes and where a fixed rule remains sufficient. Together, these results support observation preservation as a useful design choice and identify a conditional benefit of source learning for structural retrieval. The evaluated outcome is a recorded revision destination; biological ownership and proofreading productivity remain separate endpoints.

Primary area:
applications to neuroscience & cognitive science
Keywords:
connectome simulation, neural dynamics, recurrent networks, graph learning, model reduction, computational neuroscience, boundary conditions

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_yhWFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper asks whether connectome-learning pipelines should discard synaptic relations whose endpoints are not yet resolved into traced neurons. It formalizes an "observation graph" partitioning observed bodies into retained (traced) and unresolved segments, and separates an observation contrast (access to the unresolved channel, within a fixed scorer family) from a learning contrast (learned scorer vs. a fixed rule).

 The central methodological device is leave-fragment-out source supervision: count-conserving fragments sampled from MaleCNS traced owners, subtracted from the positive before feature construction, with an exact hypergeometric analysis of residual self-overlap. Models (linear and a 1,153-parameter MLP with three seeds) are frozen and evaluated on MANC v1.0→v1.2 revision-destination retrieval: 557 donor segments in 35 components, each ranked against all 23,200 earlier traced candidates, with tie-integrated, component-macro metrics. 

  The segment-aware neural scorer reaches Hit@20 of 0.5399 vs. 0.4866 (retained-only arm) and 0.4666 (fixed cosine), while Hit@1, Hit@5 and MRR all decrease. A separate masked-adjacency task improves macro AP from 0.3800 to 0.4708 with retained unresolved context. The endpoint is a recorded revision destination, not biological ownership or proofreading productivity.
Critical strengths

Full-candidate ranking against 23,200 candidates with uniform tie integration and component-macro aggregation is materially more rigorous than the sampled-negative protocols common in this area, and the support-stratified analysis (Table 2/8) honestly shows that learning adds nothing over fixed cosine for unresolved-only donors.

The leave-fragment-out construction addresses an exact supervision leak (the query retrieved through its own counts inside the positive), and the finite-population expectation analysis is the right way to show what signal survives subtraction

Critical weaknesses

The +0.0533 Hit@20 observation contrast is an average over 35 non-independent components with no uncertainty attached. Removing the five largest component differences leaves +0.0172, so a small number of components carry much of the result, and no bootstrap, permutation, or leave-one-component-out analysis is reported

Meanwhile, the paper's central contribution that is the leave-fragment-out residualized supervision is not compared empirically against naive query-with-owner supervision or partial subtraction on the target task, so the reader cannot tell whether it matters at all.

The correspondence experiments have a design problem, The model learns with the unresolved relation features present and unrestricted, and only then does the paper shuffle or zero them at test time. But shuffling everything also wrecks the candidate marginals, so when performance falls from 0.5399 to roughly 0.49, there is no way to tell how much of that drop is about correspondence and how much is plain distribution shift. A few controls would settle the question: retrain without those features in the first place, use permutations stratified by degree and mass so the marginals survive, and compare against simple degree or popularity baselines. None of them appear in the paper.

The learning story is thin for similar reasons. A small MLP gets 200 training updates on 768 tasks with three seeds, and its only competition is a fixed cosine baseline and a linear model nobody bothered to tune. There is no tuned linear model, no listwise loss, no gradient-boosted trees, no graph methods like node2vec or GNN link predictors, and no geometry-aware prior beyond the spatial centroids, which manage just 0.05 Hit@20 in Appendix E. There is also no oracle ceiling to show what the best achievable score even looks like.

Then there is the retained mass question. The 42.1% outgoing and 95.1% incoming numbers come only from MaleCNS, and nobody checked whether MANC behaves the same way. Until those missing baselines are run, much of what the paper calls learning could just be the model picking up degree and mass proxies.

Questions for authors
  1. Does leave-fragment-out supervision actually change MANC outcomes relative to naive (query-left-in) or partially subtracted supervision, on both Hit@20 and the rank-sensitive metrics?

  2. What do component-level cluster-bootstrap or permutation intervals on the +0.0533 contrast look like, and when was K=20 fixed relative to first inspection of target results? Please provide full cumulative Hit@K curves (K=1–100) with component-level bands.

  3. Does the gain survive conditional permutations that preserve candidate degree, mass, and component strata, and comparison against degree-only and popularity baselines? How do retrained-off and zeroed-at-test arms compare?

  4. Can a tuned linear model, a listwise objective, a gradient-boosted tree over the same 18 features, or a standard learned-graph baseline (e.g., node2vec or a GNN link predictor) match the MLP?

  5. Does the 42.1% & 95.1% retained-mass asymmetry hold in MANC, and can the pipeline be frozen and evaluated on a held-out component set or a different release pair to give a genuinely prospective test?

Reviewer_RDUwOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates relational learning on incomplete connectome reconstructions by formulating an observation graph that retains synaptic relations involving unresolved neuronal segments (U) alongside already traced neurons (R). The authors argue that restricting inference to traced bodies (W_RR) discards valuable observed incidence in the cross-blocks (W_RU, W_UR, W_UU), and that evaluating learned models requires separating the value of preserving unresolved observations from the value of learned feature combination. Using count-conserving leave-fragment-out supervision on MaleCNS (which subtracts sampled query counts q from the positive owner x_o prior to computing features to remove the direct self-overlap shortcut) and frozen zero-shot transfer to recorded MANC revisions (557 donor segments across 35 components ranked against 23,200 traced candidates), the authors show that a compact 18-feature neural scorer improves Hit@20 coverage and masked-adjacency average precision when unresolved context is preserved.

Critical strengths
  1. The statistical hygiene and problem formulation are rigorous, particularly the exact mathematical removal of the query self-overlap shortcut in Section 3.1, the explicit separation between observation gains and learning gains, and the analytical integration of exact score ties.

  2. The correspondence intervention (shuffling vs. zeroing unresolved relation features in Figure 2b and Table 6) and the support-stratified breakdown (Table 2) clearly isolate where the neural scorer helps, showing that learned feature combination improves Hit@20 for donors with retained support.

Critical weaknesses
  1. The results in Table 1 highlight an interesting trade-off across different ranking budgets: the learned Neural R+U scorer provides a solid boost in broader candidate retrieval (Hit@20), while the training-free Fixed Joint Cosine (R+U) baseline remains very strong at the very top of the ranked list (Hit@1 and MRR). It would be interesting to hear the authors' thoughts on what drives this trade-off between top-1 precision and top-20 recall, and whether simple adjustments such as softmax temperature scaling or hybrid scoring could combine the strengths of both.

  2. The comparison between zero-shot transfer from MaleCNS and direct cross-validation on MANC in Appendix Table 5 is a very informative diagnostic. We wonder if the authors could share some discussion on what makes direct training on the smaller MANC revision cohort more challenging, and whether sampling spatially contiguous neurite segments on the source connectome (rather than uniform synapse subsampling) might further improve transfer to real proofreading fragments.

Questions for authors

Could the authors share their intuition on why the learned models trade off Hit@1 and MRR for higher Hit@20 compared to Fixed Joint Cosine in Table 1, and whether a simple calibration or hybrid combination with cosine similarity could preserve top-1 accuracy? Additionally, I would appreciate the authors' thoughts on why direct cross-validation on MANC in Table 5 is more challenging than zero-shot transfer from MaleCNS.

Reviewer_c7yLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether preserving connectivity information from unresolved neuronal segments can improve learning on incomplete connectomes. The authors distinguish the value of additional connectivity observations from that of learned scoring and investigate these questions through two tasks. The primary task uses leave-fragment-out supervision on MaleCNS to train retrieval models, which are then transferred to MANC to predict the recorded destinations of unresolved segments. A supporting masked-adjacency task further evaluates the predictive value of unresolved connectivity information. Experimental results show improvements in Hit@20 for fragment retrieval and average precision for masked-adjacency prediction, highlighting the benefits of preserving unresolved connectivity.

Critical strengths
  • The authors try to address the challenge of utilizing connectivity information from unresolved segments in incomplete connectomes. The problem has practical significance and offers a meaningful direction for representation learning on incomplete graphs.

  • The authors evaluate the proposed approach through a primary retrieval task and a supporting link prediction task. Ablation studies and controlled comparisons further examine the contributions of additional connectivity information and learned scoring.

Critical weaknesses
  • Insufficient motivation in the Introduction. The authors introduce existing works such as NEURD and Autoproof, but do not clearly explain how current reconstruction pipelines handle unresolved segments or why some learning pipelines discard this useful connectivity information. The authors should better clarify the limitations of existing methods and the core problem addressed in this paper.
  • Unclear introduction of the research task. Key concepts such as ‘donor segment’, ‘traced candidate’, and ‘recorded destination’ are not clearly explained, making the problem setting difficult to understand. Adding an illustrative figure would help readers better understand the research problem and task setting.
  • Insufficient justification for selecting Hit@20 as the primary evaluation metric. Table 1 and Figure 2a show that the main advantage of R+U appears at Hit@20, but the authors do not sufficiently explain why K=20 is selected. In addition, only K=1, 5, and 20 are reported, leaving relatively large gaps. Results at intermediate K values would help clarify when the advantage of R+U emerges and whether it remains consistent.
  • Insufficient analysis of performance degradation. In Table 1, Neural R+U improves Hit@20 but underperforms the baselines on Hit@1, Hit@5, and MRR. Although the authors acknowledge this observation, they do not sufficiently explain why introducing unresolved connectivity improves Top-20 coverage while reducing ranking quality at smaller K values.
  • Organization of Related Work. Moving part of the related work discussion from Section 6 to the Introduction would help establish the research background and motivation more clearly.
Questions for authors
  • How do existing reconstruction and learning pipelines handle unresolved segments? What specific limitations of these approaches motivate the proposed method?
  • Why is Hit@20 selected as the primary evaluation metric? Could the authors report results at intermediate K values to show when the advantage of R+U emerges and whether it remains consistent?
  • Why does Neural R+U improve Hit@20 while reducing Hit@1, Hit@5, and MRR? Could the authors provide further analysis of this ranking trade-off?
Reviewer_EZnTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates whether unresolved segments in connectome data carry useful information. The main experiment treats unresolved segments as queries, constructs features from their connection profiles to retained neurons (R channel) and to other unresolved segments (U channel), and evaluates retrieval hit rates over all traced neuron candidates. Results show that adding the U channel improves Hit@20 while slightly decreasing Hit@1, Hit@5, and MRR.

Critical strengths

The paper designs targeted experiments to prevent self-overlap leakage. The experimental protocol and feature selection are described clearly and completely.

Critical weaknesses
  1. The significance of this problem is not entirely clear to me. Unresolved segments naturally contain extra information, so retaining them should be more informative than discarding them. The real challenge lies in how to leverage them effectively. However, the experiments rely on the complete retained graph from MANC v1.2, rather than showcasing novel applications that unresolved segments alone could enable. In my view, the central goal of connectome reconstruction is to improve its efficiency and coverage, not to focus on exploiting unresolved segments in isolation.
  2. The topic has limited alignment with ICLR. Its methodology largely consists of a procedural experimental protocol and basic graph-theoretic measures. The only machine learning component is a simple MLP used to score retrieval results. The empirical outcomes are modest: an improvement in Hit@20 comes at the cost of degraded performance on other metrics. Although the work may have merit, it appears better suited for a biomedical engineering venue.
Questions for authors
  1. Without a retained graph serving as ground truth, what practical applications can unresolved segments offer?
  2. In Table 1, why does the R+U setting perform worse than U on Hit@1 and MRR? Does this indicate that unresolved segments introduce noise?
  3. The experiments rely on handcrafted features. Could these features be insufficient, thereby weakening the persuasiveness of the retrieval accuracy comparisons?
  4. Overall, why do the Linear and Neural models underperform the Fixed Joint Cosine baseline (Table 1) ?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.