acceptodds
Under review as a conference paper at ICLR 2027

SPIRE: Shared-Private Representation Learning for Continuous Multi-Region LFP under Perturbation

Abstract

Learning shared and region-specific representations from continuous multi-region neural signals is challenging when observations are nonlinear, autocorrelated, spectrally structured, and temporally misaligned. These challenges are particularly pronounced in local field potentials (LFPs), where separating activity that is coordinated across regions from activity specific to individual regions is important for studying distributed neural systems and their response to intervention. We introduce **SPIRE** (Shared-Private Inter-Regional Encoder), a nonlinear, time-resolved framework for learning shared and private representations from continuous multi-region LFP. SPIRE combines region-specific sequence encoders with constrained temporal correspondence, shared-only reconstruction, and variance regularization for shared/private separation. On LFP-oriented synthetic benchmarks with known ground-truth factors, SPIRE preferentially assigns shared and private information to the intended latent spaces under nonlinear observation mappings, time-varying delays, and spectral confounding. On held-out human GPi-STN LFP, nonlinear prediction reveals strong transfer of shared representations across regions, whereas cross-region prediction from private representations remains at a structured temporal null, supporting functional rather than strict statistical separation. We then freeze representations learned from off-stimulation recordings and apply them to intracranial deep brain stimulation (DBS) data. Stimulation induces larger changes and carries more condition information in shared than private representations. A matched CTAE analysis reproduces the qualitative shared-over-private effect while retaining substantially less stimulation information. Together, these results establish SPIRE as a shared/private representation-learning framework for continuous multi-region LFP and show that DBS-related changes are preferentially expressed in representations shared across the recorded regions.

Primary area:
applications to neuroscience & cognitive science
Keywords:
representation learning, multi-region neural data, shared-private representations, local field potentials, neural time series, deep brain stimulation, disentanglement

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_LeWSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SPIRE, a method that decomposes continuous multi-region local field potentials (LFPs) into shared (cross-regional) and private (region-specific) latent trajectories. SPIRE employs a separate GRU encoder for each region, whose latent representation is split into shared and private components: the shared component is designed to capture information in brain activity that is transferable across regions, while the private component captures region-specific information. To induce this separation, the shared latents of different regions are trained to align with one another, the shared and private latents within each region are trained to be de-correlated, and an additional objective reconstructs the signal of another region from the shared latents alone.

The authors first evaluate SPIRE on synthetic LFP data with known ground-truth factors and report that SPIRE outperforms the baselines under each nonlinear mixing, time-varying inter-regional delays, and spectral confounding condition. They then apply SPIRE to real GPi-STN recordings from pediatric dystonia patients. Because the true shared and private factors are unknown in real data, they validate the model by testing whether the latents of one region can predict those of another, and show that shared latents support cross-regional prediction beyond a structured temporal null, whereas private latents do not. Finally, the authors train SPIRE only on stimulation-free data and apply it to recordings acquired during deep brain stimulation (DBS). In this experiment, stimulation shifts the shared latents more than the private latents, and the stimulation frequency is also decoded more accurately from the shared latents. Based on these results, the authors argue that DBS effects are expressed more strongly in shared brain activity than in region-specific activity.

Critical strengths

A well-designed synthetic data experiment : To evalute that the shared and private factors are actually separated and assigned to the shared and private latent components as intended, the authors construct synthetic LFP data with known ground truth (known shared and private factors), train the model and evaluate how well the learned latents predict each intended factor using a 2 ×\times 2 recovery matrix (SS, SP, PP, PS) as the evaluation metric. Specifically, they assess whether the shared latents predict the shared factors well (high SS) but not the private factors (low SP), and whether the private latents likewise exhibit high PP and low PS. In particular, the synthetic data span a range of settings, from a relatively easy linear setting to settings that incorporate characteristics of real LFP data, such as nonlinear mixing, time-varying inter-regional delays, and spectral confounding. These results show that SPIRE performs better than the baselines in the more challenging settings, implying evidence that SPIRE could achieve the intended separation training effects even under LFP-like data settings. Moreover, since the shared and private factors are generally not clearly known in real LFP data, directly verifying whether the intended separation occurs on synthetic data that reflect the properties of LFPs makes this a well-designed experiment and I think this is a key experiment which supports the strengths of SPIRE.

Critical weaknesses

Limited methodological novelty : As the authors themselves acknowledge, learning shared and private representations with separate encoders while encouraging joint reconstruction, shared-private decorrelation, and alignment among shared representations are well-established approaches in representation learning, and SPIRE appears to be a framework that appropriately combines them, which leaves its methodological novelty considerably limited. For example, CTAE [1], one of the main baselines in the paper, already learns region-wise shared and private latent trajectories from multi-region electrophysiology data, and already proposes encouraging alignment among shared latents, reconstruction from shared latents alone, and shared-private decorrelation. Therefore, SPIRE does not appear to make a new methodological contribution over CTAE, but rather to replace some of its components (using a GRU encoder instead of a transformer, or using VICReg [2] for aligning the shared latents). ConvAlign addresses inter-regional temporal misalignment, but the problem of inter-regional delays has already been addressed in prior work, such as DLAG [3], the other baseline, and it remains unclear what new methodological insight ConvAlign provides beyond existing approaches to resolve temporal misalignments. Since the core components of the SPIRE framework thus appear to be largely covered by prior work, in my opinion, the authors should provide stronger justification for the methodological novelty of their approach. I believe this is the most critical point the authors needs to address and I consider this weakness as the most consequential.


 [1] Sristi et al., Coupled Transformer Autoencoder for Disentangling Multi-Region Neural Latent Dynamics, 2025.
 [2] Bardes et al., VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning, 2021.
 [3] Gokcen et al., Disentangling the flow of signals between populations of neurons, 2022.

Questions for authors
  1. Methodological novelty : Could the authors justify the core methodological novelty or contributions of SPIRE? If the methodological contributions of SPIRE’s key design choices can be justified, it would help in reassessing the paper’s core contribution. As mentioned earlier, I think this is the critical point, so I'm very open to discuss it further with the authors.

  2. More extensive empirical validation : Even if the methodological novelty is limited, the contribution within the LFP domain could be recognized if it were validated that SPIRE’s design choices, which combine several components for the LFP data domain, yield clear performance gains. However, the current level of empirical validation seems insufficient to support such a claim. First, only CTAE and DLAG are presented as baselines, and given that CTAE shows low recovery even in the approximately linear D0 setting, I wonder whether the baselines were tuned sufficiently and fairly. Although the appendix describes tuning configurations, could the authors provide further clarification on whether the baselines were tuned enough? If so, could the authors explain the low performance of CTAE even in the approximately linear D0 setting? Also, I would like to know whether SPIRE’s performance advantages in the LFP domain can be clearly demonstrated against a broader set of sufficiently strong baselines, such as other multi-view models. In addition, just as the authors constructed separate synthetic data to verify this separation, are there real datasets in other domains that would allow the effect of this separation to be validated more directly, and if so, could such experiments be included? If these points are addressed with more extensive empirical validation, the paper’s contribution could be recognized even if its methodological novelty is limited.

  3. Neuroscientific interpretation of the DBS finding : In the DBS experiment, the authors show that stimulation shifts the shared representation more than the private representation and that the stimulation frequency is decoded more accurately from the shared representation. Is it possible to conduct analyses that provide deeper neuroscientific understandings of these results, beyond the interpretation that DBS acts across multiple regions? Alternatively, beyond DBS, are there other experiments in which the separation of shared and private factors yields results consistent with established neuroscientific findings? If such experiments and analyses enable a deeper neuroscientific understanding, the paper’s contribution would be substantially strengthened.

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

This paper introduces SPIRE, a framework that separates recordings from multiple brain regions into shared and region-specific representations. It validates this separation using synthetic signals and human local field potential recordings, then applies the learned representations to deep brain stimulation data. The results suggest that stimulation-related changes are more strongly expressed in shared activity across regions than in region-specific activity.

Critical strengths
  • Synthetic experiments and held-out human-data analyses both provide complementary experimental results that could show whether SPIRE captures the intended types of information.

  • Applying representations learned without stimulation labels to human DBS recordings demonstrates practical value, while extensive ablations and sensitivity analyses strengthen confidence in the findings.

Critical weaknesses
  • The baselines used for the experiments seem insufficient. CTAE is the only nonlinear neural baseline, and it shows weak recovery and specificity performance even in the synthetic setting. I currently cannot tell from this paper if this is due to the authors not properly implementing/training the CTAE model, or the CTAE model is genuinly inable to properly function for LFP. CTAE largely fails even for the D0 regime, which warrants a close investigation about the cause of this failure. It could be suboptimal optimization, insufficient data capacity, or an inherent problem with the CTAE such as its decomposition objective. Considering CTAE was the only baseline for the real human experiments, this point needs to be addressed.

  • The DBS findings need stronger controls for latent geometry. Shared and private latents receive different variance constraints, so larger centroid displacement in raw latent coordinates does not by itself establish greater perturbation sensitivity. The displacement comparison should use normalized or whitened coordinates, with dimensionality accounted for. Additionally, additional details are needed to properly understand and analyze the experiments: currently it is unclear whether decoding holds out complete stimulation trains or recording blocks. Without these details and related analysis, it is difficult to properly judge the reported DBS results.

Questions for authors
  • What explains CTAE’s weak recovery and specificity, particularly in the simple D0 regime? Can the authors provide convergence and hyperparameter-sensitivity results to distinguish training or capacity limitations from a genuine limitation of CTAE, and provide some results from another relevant neural baseline if they could afford to run additional experiments?

  • Does the shared-over-private displacement effect persist after accounting for latent scale and dimensionality? Please report the comparison using normalized or whitened representations.

  • How are the training and test sets constructed for DBS decoding? Are complete stimulation trains or recording blocks held out, and does the decoding advantage persist under this split to limit leakage from temporally related windows?

1 comment
Reviewer_DZcGOpenReview ↗

Response to author comments

I thank the authors for their detailed response. My concerns regarding CTAE’s failure remain unresolved, as I still cannot determine whether it reflects a limitation of CTAE or implementation or training issues made by the authors of this paper. However, the additional baselines help address my main concern about the lack of a strong nonlinear baseline. The authors have also partly addressed my other concerns, and I understand that completing all the requested experiments may be difficult within the limited discussion period. Taking these improvements into account, I will raise my score to 3.

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

This paper introduces SPIRE, a framework for learning time-resolved shared and private representations from continuous multi-region local field potentials. The method combines region-specific recurrent encoders and decoders with constrained temporal alignment, shared-only reconstruction objectives, and covariance and variance regularization. Evaluation includes synthetic benchmarks with nonlinear mixing, time-varying delays, and spectral confounding, followed by analyses of human GPi-STN recordings. Models trained exclusively on off-stimulation data are frozen before analyzing DBS-related changes. The results show preferential recovery of the intended synthetic factors, stronger cross-region predictability between shared representations, and greater stimulation-related displacement and decoding accuracy in shared than private representations. Additional experiments examine component ablations, robustness to modeling choices, a three-region extension, and a preliminary application to rat CA1-PFC recordings.

Critical strengths
  1. The paper addresses a meaningful representation-learning problem with evaluation tailored to the characteristics of continuous neural recordings. Nonlinear observation mappings, temporal misalignment, and spectral confounding are explicitly incorporated into the synthetic benchmarks. The D3 setting is particularly informative because it tests whether the model can recover shared factors even when stronger, slower activity is private.

  2. The synthetic evaluation goes beyond reconstruction quality or intended-factor recovery. Reporting the complete shared/private recovery matrix and specificity margins directly examines whether information is preferentially assigned to the appropriate branch. SPIRE achieves positive specificity margins in both branches across all four regimes, while the stronger performance of DLAG in the linear control provides a useful qualification of SPIRE's advantages.

  3. The stimulation analysis uses a sensible separation between representation learning and downstream evaluation. Training and model selection rely exclusively on off-stimulation recordings, reducing the risk of shaping the representation around stimulation labels. The matched CTAE comparison also provides useful corroboration: both methods show stronger stimulation decoding from shared representations, while SPIRE achieves substantially higher accuracy, including 67.7% versus 38.4% for the shared representations in Table 4.

  4. The supplementary evaluation is substantial, including loss-family and training-control ablations, no-lag retraining, dimensionality sensitivity, an alternative reconstruction objective, and checks of representation stability under stimulation. I also appreciate the explicit distinction between functional shared/private organization and statistical independence or causal communication. The three-region and four-animal experiments provide encouraging initial evidence beyond the main two-region analysis, with appropriately qualified claims.

Critical weaknesses
  1. The methodological contribution would benefit from sharper attribution. SPIRE combines several established ideas, and CTAE already addresses a closely related shared/private representation problem. The ablations in Appendix A.2.5 are useful, but primarily evaluate intended-factor recovery using CCA. Extending these ablations to the complete nonlinear recovery matrix would clarify which additions improve branch specificity, rather than only improving recovery of the intended factors. This would better connect the component-level evidence to the paper's central contribution.

  2. The human-data validation does not yet report the complete cross-region prediction matrix. Table 3 includes shared-to-shared and private-to-private prediction across regions, whereas the mixed shared/private predictions are within-region. Given the substantial within-region nonlinear dependence, weak private-to-private transfer alone does not establish that private representations contain little information about the other region's shared representation. Reporting both directions of cross-region private-to-shared and shared-to-private prediction would strengthen the operational interpretation. This does not require strict statistical independence, which the paper appropriately does not claim.

  3. Some evaluation and statistical details need clarification. Appendix A.1.2 describes an 80/20 training/validation split, whereas Appendix A.3.1 describes an 80/20 training/test split; the relationship between these splits and latent-dimension selection should be stated explicitly. Holding out complete 0.5-second windows is preferable to splitting individual timepoints, but the manuscript should also explain whether temporally adjacent windows or windows from the same stimulation episode can appear in different folds, particularly for stimulation decoding. Finally, because the human analyses include multiple hemispheres from the same patients, the treatment of within-patient dependence in confidence intervals and significance tests should be made explicit.

Overall, these concerns limit the strength of the methodological attribution and some aspects of the empirical interpretation, but they do not outweigh the paper's useful formulation and substantial supporting evaluation. My assessment is mildly positive.

Questions for authors
  1. Could the authors specify the complete training, validation, and test hierarchy for representation learning, dimensionality selection, and downstream decoding? Are decoding folds grouped by complete stimulation episodes or temporally separated recording blocks?

  2. What are the nonlinear prediction results for private GPi to shared STN, private STN to shared GPi, and the corresponding reverse directions? Do these additional comparisons support the same functional shared/private interpretation?

  3. Could the main component ablations be evaluated using the complete nonlinear recovery matrix and specificity margins from Tables 1–2? Which components contribute most consistently to reducing crossed recovery?

  4. How is off-to-on centroid displacement normalized for latent dimensionality and baseline variability? Also, are the main significance tests and confidence intervals robust to treating patients, rather than hemispheres alone, as the independent sampling units?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal revision submitted

Dear reviewers and Area Chair,

We have uploaded a revised manuscript that incorporates the new analyses and clarifications from the rebuttal. To make the changes easy to find, here are the main updates relative to the original submission.

  • §4 and Appendices A.2.4, A.2.6, and A.2.8. We added the Shared-AE and MR-LFADS comparisons, additional CTAE tuning and a positive control, and experiments that distinguish the contributions of the alignment loss and the temporal aligner.
  • §5.1 and Appendix A.5.1. We expanded Table 3 to show all four shared/private prediction pairs in both GPi–STN directions and added patient-level analyses. Shared-to-shared prediction was strongest in both directions, but some mixed predictions were above the temporal null.
  • §6 and Appendices A.7–A.8. We clarified the original pointwise decoding protocol and added the whole-window holdout analysis for SPIRE and the dimension-normalized displacement analysis. The matched CTAE comparison remains in the main text and is explicitly labeled as pointwise.

Thank you for the suggestions that motivated these additions. We’d welcome any feedback on whether the revision addresses your main concerns or whether anything still needs clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.