acceptodds
Under review as a conference paper at ICLR 2027

Learning Multimodal Population Relations for Cross-recording Neuron Matching

Abstract

Establishing reliable neuronal correspondences across recordings is essential for comparing neural dynamics across individuals and linking functional activity to circuit organization. However, this remains challenging because neuronal identities cannot be reliably determined from individual features alone, as spatial organization and functional activity patterns vary across recordings. Here, we propose NeuReL, a supervised neuron matching framework that integrates multimodal population relations by jointly learning positional and functional relationships among neurons within their surrounding populations. NeuReL first captures population-level representations within each recording and then refines correspondences by enforcing relational consistency across recordings. On held-out animals, NeuReL achieves significant improvements over the strongest evaluated baselines across multiple C. elegans datasets. We further validate NeuReL on longitudinal neuron matching in zebrafish, demonstrating its applicability across species and recording conditions. These results demonstrate that multimodal population relations provide a powerful basis for cross-recording neuron matching, enabling reliable neuronal correspondences for comparing neural activity across experimental sessions and individuals.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Multimodal representation learning, relational representation learning, optimal transport, population-aware representation learning, computational neuroscience, neural data analysis

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_6FHWOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces NeuReL, a framework for neuron identity matching across activity records. NeuReL jointly uses geometric and activity relations to build population relation embeddings, and learns individual representations for each neuron. The matching score is computed by embedding similarity and temporal relation correction process. Experiments show that leveraging geometric and activity signals together improves matching accuracy over baselines using only geometric or activity signals.

Critical strengths

1.The problem is important in neuroscience and remains underexplored from a ML perspective. The authors evaluate under cross-animal matching for C.elegans and longitudinal matching for zebrafish, demonstrating the applicability of the method under distinct settings. 2.The idea of using population information for identifying the neuron is well-motivated, and embedding the geometric and activity relation as relation-conditioned attention is intuitive.  3.The authors compare against geometry-only and activity-only methods, as well as conduct correlation analysis and case studies, supporting the claim that activity and geometry information are complementary.

Critical weaknesses

1.Supervised learning requires massive human-labeled datasets, and the annotation may become noisier as we scale to animals with more neurons. However, the dependence on annotation quantity and quality is not sufficiently characterized in the paper. Controlled experiments on reduced and corrupted annotations could help to better understand the method’s practical reliability. 2.Missing atlas relations are treated as observed zero-valued relations. For example, there is a considerable portion of directed atlas entries on Kato that correspond to identity pairs that were never jointly annotated. These entries are assigned to zero but still contribute to the relation-discrepancy penalty. This conflates missing information with an observed relation and may penalize otherwise correct matches.

Questions for authors

1.For the C. elegans datasets, where canonical neuronal identities are available, how would NeuReL compare with a direct identity classifier using the same multimodal population encoder, followed by Hungarian assignment to enforce one-to-one correspondence? 2.How does performance change as the fraction of annotated training neurons decreases or incorrect correspondence labels are introduced? 3.Could the authors perform an ablation by masking out the unannotated pairs (see W2) and test whether the current zero-vector treatment introduces additional error?

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

Neural Relational Learning (NeuReL) combines neuronal geometry and neural activity through multimodal population relations. A relation-conditioned Transformer and relation-aware neuron matching use population context to match neurons across recordings. Evaluation covers two C. elegans datasets and longitudinal recordings from held-out zebrafish.

Critical strengths

1.The problem is important, and the multimodal relational formulation is well motivated. The Top-1 improvements on Atanas and Kato are large, with ablations supporting the contribution of activity and population context.

2.Animal-level splits and the detailed reference-choice, rejection, and fixed-encoder analyses make the evaluation reasonably transparent.

Critical weaknesses

1.The main evaluation only partially tests partial matching. Hungarian decoding first selects ground-truth-eligible query rows and excludes the unmatched state. This removes unknown queries before assignment, so the reported score does not establish performance when all neurons must be decoded together. Top-1 is a separate result.

2.Population-atlas construction is fragile under missing annotations. Missing relation entries are filled with zeros, including 55.97% of entries on Kato. The single medoid also outperforms the population atlas there, 86.43% versus 74.62% Top-1. This weakens the case for the chosen atlas representation.

3.The smaller zebrafish gain needs uncertainty at the fish level. The 0.94-point Top-1 improvement is summarized over 101 pairs from eight fish. Pair-level standard deviations do not establish whether this gain is consistent across fish; training-seed variability is also insufficiently characterized.

Questions for authors

1.What are the Hungarian and unmatched-rejection results when assignment uses all query neurons and annotation-based filtering is applied only to scoring?

2.How do the existing results change when missing atlas relations are masked, and when zebrafish differences are summarized per fish?

1 comment

Thank you for your clarification and additional experiments. I will maintain the original rating.

Reviewer_DqMKOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present a novel method for supervised cross-neuron matching (across animals or time), leveraging both inter-neuron relationships based on position and activity. In my view, this work has the potential to be a strong paper, as the method itself seems strong, but the value of the method isn’t sufficiently clear in the current version, and I have potential concerns about robustness to new datasets.

Critical strengths

-Strong performance relative to existing approaches

-Comparisons to many baselines

-Demonstration of the method on multiple real datasets

Critical weaknesses
  1. The importance and use cases of this method aren’t currently clear. If all neurons can be labeled (which is necessary for this supervised approach to work), when would you need to do cross-neuron matching?
  2. There are many knobs that can be tuned in the architecture and optimization (e.g. tables 9 and 10). It is currently unclear whether these knobs have just been tuned to work on the existing datasets, and whether this method would work out of the box on a new dataset, or if finetuning would be needed (and if finetuning is needed, how would this be done on held-out data)?
  3. The framing has a bit of overclaiming - this neuron matching problem is not really a problem across all of neuroscience, since in the majority of animal systems, individual neurons aren’t cared about. C elegans is indeed a common example where this is cared about, and there are other examples as well, like the specific zebrafish spinal segments (in fact, this is not something that would be done in an entire zebrafish brain, which the abstract would make a reader believe). Neuron matching is an important problem, I would just frame it more accurately and not overclaim the broadness of its applicability.
Questions for authors

-Could you clarify what baseline methods are supervised versus self-supervised/unsupervised? It’s currently hard to know whether these are apples-to-apples comparisons.

-What are the data regimes in which your method works well/ doesn’t work well? It would be useful to investigate this on either real or synthetic data, so users have confidence on when to use this method.

-In Fig 2B, could you walk through an example neuron? (This would be helpful in the text). I’m confused because I would have assumed that the incorrect predictions were for spatial hard candidates.

-Can you clarify if neural activity needs to come from the same behavior across animals?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.