acceptodds
Under review as a conference paper at ICLR 2027

RAVEL: Factorizing multi-region neural communication into reusable propagation routes and dynamic recruitment

Abstract

Inter-regional neural interactions can change rapidly with behavior and internal state, but such changes are mechanistically ambiguous: they may reflect changing transmission properties, or changing recruitment of otherwise stable communication mechanisms. Existing multi-region models capture delayed and time-varying interactions, but do not explicitly separate these alternatives. Motivated by parallel biological communication routes and their state-dependent engagement, we introduce RAVEL, a conditional spectral model for directed communication in multi-region local field potential (LFP) recordings. RAVEL predicts target spectral activity from a source region and factorizes the source-to-target transfer into reusable delayed propagation components, with explicit transport delays and frequency-dependent filtering, and time-varying recruitment that determines when each component is expressed. In synthetic experiments, RAVEL recovers stable propagation physics and changing recruitment when that mechanism generates the data, while a matched dynamic-delay model is favored when transmission delays genuinely change. On Allen Brain Observatory LFP recordings, propagation dictionaries are reproducible across mice, transfer across behavioral contexts with little loss in held-out predictive performance, and exhibit behavior-dependent recruitment. Locally learned transfers can also be composed to predict a withheld long-range interaction without fitting the endpoint directly. RAVEL further performs competitively on complete-target cross-region prediction against established models. Together, these results support separating relatively stable propagation structure from its dynamic recruitment as an interpretable description of multi-region neural communication.

Primary area:
applications to neuroscience & cognitive science
Keywords:
multiregion neural communication, neural dynamics, spectral modeling, compositional generalization, cross-spectral analysis, transfer learning, interpretable machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Z84XOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed a framework, RAVEL, which models time-varying neural communication as the dynamic recruitment of fixed propagation components. The proposed model is evaluated on both synthetic and real mouse electrophysiological recording datasets. The results tries to domanstrate parameter recovery, cross-animal transfer and neural activity prediction.

Critical strengths
  1. The paper provides an interesting perspective of time-varying effective delays in neural communication.
  2. The decomposition into fixed propagation components and dynamic recruitment coefficients is intuitive reasonable and potentially useful in further understanding of neural communication mechnism.
  3. The experimental results cover boardly with both synthetic and real mouse electrophysiological recording datasets.
Critical weaknesses
  1. Lack of comparison of existing LTI methods. Although the proposed method might be a good option for understanding the neural communication mechnism, the advantages over general LTI methods are not clearly demonstrated.
  2. Insufficient justification of the proposed factorization of the proposed propagation. The comparisons in the paper mainly focus on static-recruitment and dynamic-delay variant, but it lacks a matched capacity comparison with general LTV models using learnable LTI components. Therefore, it remains unclear whether the reported benefits caused by the specific propagation parameterizaiton or simply from dynamic basis recruitment.
  3. Limited evidence. In the synthetic experiments, RAVEL achieves an NMSE of 0.212, compared with 0.217 for the dynamic-delay model under fixed-delay generation. Since the difference is really small, it does not establish that the proposed method can reliably outperform alternative mechanisms. While the cross-animal experiments show similar held-out NLL gains for pooled and independently fitted dictionaries (0.025 vs. 0.028), but the spatial coupling and recruitment parameters remain recording-specific. It could support the predictive ability, but could not sufficiently support the inferred components correspond to conserved propagation mechanisms.
  4. Statistical evidence to support the claims of generalization ability is not enough. For example, for VISam←VISp, the improvement over MRM-GP is not statistically significant after correction (p=0.1406). And the multi-hop experiment reports a relatively modest spectral residual R square equals to 0.172 and is evaluated on only 36 held-out events from a single recording session.
Questions for authors
  1. Could the authors compare the proposed framework against existing LTV model with general learnable LTI components?
  2. Could the authors provide more evidence to support the proposed propagation factorization shows stable mechanisms shared across animals, given to the similar predictive performance between the pooled and independently fitted dictionaries?
  3. Could the authors further discuss the generalizability of the multi-hop results, given the limited evaluation across recording sessions, and the relatively modest improvements over MRM-GP?
Reviewer_2pZJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose that changing interactions between brain regions can be explained by several components with fixed delays whose strengths vary over time. This is an interesting approach. The transfer across mice and behavioral contexts, and the prediction of long-range interactions by combining locally fitted models, are encouraging. However, the biological interpretation would benefit from stronger validation.

Critical strengths
  1. The authors propose that changing inter-regional interactions can arise from fixed communication components with changing strengths. This is an interesting idea and provides a useful way to separate relatively stable interactions from their behavioral modulation.
  2. The learned components retain predictive value across mice and between active and passive conditions. The model also performs competitively with existing methods.
  3. Locally fitted interactions can be combined to predict a withheld long-range interaction. This is a useful demonstration that the model can make predictions beyond individual area pairs.
Critical weaknesses
  1. While the idea is interesting, the experimental results do not clearly demonstrate that the observed interactions follow the proposed mechanism. The direct comparison between changing component strengths and changing delays is mainly based on simulations.
  2. The biological meaning of the components remains unclear. Shared inputs and local population dynamics could also contribute to the delayed LFP relationships. Predictive performance alone cannot distinguish these possibilities.
  3. The hit–miss difference is potentially interesting, but remains descriptive. It is unclear whether this difference is consistent across mice or reflects differences in movement, arousal, or licking.
  4. The experimental validation covers a small number of area pairs, and the long-range composition result comes from one mouse. It is therefore difficult to assess how general these findings are.
  5. The reproducibility of individual components needs clarification. Some delays vary considerably across mice, and the four-component model was retained despite not meeting the stated stability criterion.
Questions for authors

Suggestions:

  1. Quantify the hit–miss difference across mice and account for running, pupil size, and licking. Please clarify whether the difference occurs before or after the behavioral response.
  2. Test additional area pairs and both directions, and replicate the long-range composition analysis across mice. These analyses could be performed using the existing dataset.
  3. Clarify the choice of component number and show whether the learned delays and main conclusions remain consistent across trial subsets and different model sizes.
Reviewer_FWtkOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes that inter-regional communication in LFP recordings can be described by a small set of fixed transmission delays between region pairs, each of whose strength changes over time. This contrasts with the view that the delays themselves change over time, and suggests there could be a basis of delay components shared across subjects. To test this, the authors use a synthetic testbed, as well as real data from mice performing a visual change-detection task. The paper suggests that there exists a single shared set of 4 delays that predicts target-region LFP from the source region as well as subject-specific delays do. It further claims that these delays carry over from active to passive viewing without loss in prediction, that differences between hit and miss trials show up in how strongly each delay component is weighted over time rather than in the delays themselves, that separately fit region pairs can be chained to predict activity in a distant region, and that the model is competitive with existing multi-region models at cross-region prediction while being much faster to fit.

Critical strengths

(S1) The central hypothesis is interesting and offers a useful mechanistic perspective on inter-regional communication. If true, it could have meaningful implications for how we interpret changes in functional coupling.

(S2) The most interesting result to me is that a single shared set of 4 learned delays predicts held-out activity as well as subject-specific fits do, and that the independently fit delays roughly cluster around the shared ones (at least 3 of them, Motif 2 seems off). This supports the claim that delay components are reusable across subjects.

Critical weaknesses

(W1) There is limited evidence based on their experimental setup for the proposed claim. The synthetic testbed is largely tautological, since each model recovers data generated under its own assumptions. For example, RAVEL is expected to better explain the data model that is based on its own inductive bias. Also, given that predictive performance is equal for the two models in the fixed-delay world, there is no basis to establish predictive performance as a metric supporting the central claim. But in the real datasets, this is the central metric that is used to make the claims. To me, predictive performance cannot describe the proposed mechanism. For example, in the changing-delay world, RAVEL is the wrong model, yet it still substantially outperforms the nuisance model. So beating the nuisance model on real data, which is the main evidence in Section 3.2, is what we would expect to see even if the delays themselves change over time. The changing-delay model is never fit to the real data, so the central claim is never actually tested there.

(W2) I don’t think there is enough data to qualify the scope of the claims. Section 3.2 uses a single dataset with only 25 mice, and only a single region pair. Section 3.3 uses a single mouse with 36 test events. Section 3.4 uses 8 sessions, but with only 10 training and 5 test trials each at a single stimulus orientation, and RAVEL is significantly better in only one of four comparisons. The scope of the results is hence fundamentally limited.

Questions for authors

(Q1) Can the authors fit the dynamic-delay model, and ideally a model where both delays and weights change, on the VBN data and compare it to RAVEL? Given the synthetic results, a tie in predictive performance would be uninformative. What criterion would the authors use to decide between the two mechanisms on real data? For example, consistency of the inferred delays across mice, or transfer across contexts.

(Q2) Was the split-half reliability computed within behavior type or across all trials? Can the authors show split-half reliability within hits and within misses separately, and also compare the hit-miss difference to a null with shuffled labels and matched trial counts? Without these, it’s unclear whether hit and miss differ any more than two random subsets of trials would.

(Q3) Could the changing weights reflect changing signal quality? Source signal-to-noise likely changes around the image change. When the source is noisy, least-squares fits are biased toward smaller weights. Is it possible to rule out that the time-varying weights simply track changes in source SNR rather than changes in communication?

I think in general there is also a lot of room for improvement in the presentation of the work. For one, several terms are difficult to follow, and I think it’s worth either defining them more clearly toward the beginning or using clearer wording. Phrases such as “communication subspaces,” “recruitment profiles,” “reusable atoms,” and “nuisance and propagation residue matrices” are overly complex, and I don’t think they communicate the main ideas effectively. Also, a suggestion: move Section 2.3 closer to Eq. 1, since it defines the U and V used there and otherwise feels out of place after 2.2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.