Abstract

Electroencephalography (EEG) plays an important role in brain–computer interfaces (BCIs) and clinical neuroscience by enabling non-invasive measurement of brain dynamics. Covariance descriptors are widely used in EEG decoding for their noise robustness and spatiotemporal statistical representation. Full-rank correlation matrices have emerged as a scale-invariant alternative, but current normalization layers do not account for their intrinsic geometry. In this paper, we propose Correlation Batch Normalization (CorBN), a normalization layer tailored to full-rank correlation matrices in EEG decoding. CorBN leverages the Lie group structure of the correlation manifold to perform geometrically consistent normalization while preserving manifold structures. Our approach ensures more stabilized training of deep neural networks, while respecting the intrinsic geometry of correlation matrices. We conduct extensive experiments on three EEG decoding tasks, and further validate CorBN on a large-scale clinical EEG benchmark. The results show that integrating CorBN into correlation-based networks consistently improves accuracy and accelerates convergence compared with strong baselines. The code can be found at https://github.com/JiaruiYang-AI/CorBN.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1392_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/JiaruiYang-AI/CorBN

Link to the Dataset(s)

BCIC-IV-2a dataset: https://www.bbci.de/competition/iv/ MAMEM-SSVEP-II dataset: https://www.mamem.eu/results/datasets/ BCI-ERN:dataset: https://www.kaggle.com/competitions/inria-bci-challenge/data

BibTex

@InProceedings{YanJia_Riemannian_MICCAI2026,
        author = { Yang, Jiarui AND Hu, Chen AND Wang, Rui AND Xu, Tianyang AND Hu, Cong AND Zhou, Tao AND Wu, Xiao-Jun},
        title = { { Riemannian Batch Normalization on Correlation Manifolds for EEG Decoding } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16894},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes Correlation Batch Normalization (CorBN), a normalization layer specifically designed for full-rank correlation matrices in EEG decoding. By leveraging the Lie group structure of the correlation manifold under the Off-Log Metric, CorBN performs geometrically consistent centering, scaling, and biasing directly on the manifold, enabling stable and efficient training of deep networks while preserving the intrinsic structure of correlation-based representations.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Well-defined and technically sound problem formulation The paper identifies a clear gap: existing batch normalization methods cannot be directly applied to correlation matrices due to the unit-diagonal constraint defining a strict submanifold. Addressing normalization on this specific manifold is a meaningful and well-motivated problem in Riemannian deep learning for EEG.

    2.Principled geometric design leveraging Lie group structure CorBN is carefully derived using the Lie group structure of the correlation manifold under the Off-Log Metric. The method performs centering via Fréchet mean, scaling in the tangent space, and biasing through group action, providing a geometrically consistent analogue of Euclidean batch normalization. This principled formulation is both mathematically elegant and conceptually clear.

    3.Closed-form and computationally efficient implementation Unlike some Riemannian normalization approaches that rely on iterative procedures (e. g. , Karcher mean), CorBN admits closed-form expressions for its key operations. This improves computational efficiency and makes the method more practical for integration into deep learning pipelines.

    4.Theoretical justification of statistical behavior on manifolds The paper provides a statistical interpretation of CorBN through Riemannian Gaussian distributions, showing how the method controls mean and variance on the manifold. This strengthens the contribution beyond a heuristic design and connects it to established probabilistic principles.

    5.Consistent improvements and improved training stability Experimental results across multiple EEG benchmarks demonstrate consistent performance gains and reduced variance. Additionally, training dynamics analysis shows faster convergence and more stable gradients, which aligns with the intended purpose of normalization layers.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.Limited methodological novelty relative to prior work While CorBN is well-formulated, it is conceptually an extension of existing normalization techniques to a specific manifold. Riemannian batch normalization for SPD matrices has already been introduced, and Lie group-based normalization methods have also been explored. The main contribution here is adapting these ideas to the correlation manifold under the Off-Log Metric, which is a meaningful but incremental extension rather than a fundamentally new paradigm.

    2.Modest empirical improvements The reported performance gains are relatively small (typically around 0.5 ~ 1% across tasks), as shown in the comparison tables. While consistent, these improvements are limited in magnitude and may fall within the variability of training, raising questions about the practical impact of the method in real-world applications.

    3.Limited experimental scope and dataset diversity The evaluation is conducted on a small number of standard EEG benchmarks (e. g. , BCIC-IV-2a, MAMEM-SSVEP-II, ERN), which are relatively small and commonly used datasets. There is no validation on larger-scale or more clinically relevant datasets, which limits the generalizability and real-world applicability of the findings.

    4.Lack of strong comparisons with recent state-of-the-art methods Although several baselines are included, the comparison does not fully cover the most recent advances in EEG decoding or manifold-based deep learning. In particular, newer architectures or normalization techniques are not extensively benchmarked, which weakens the claim of state-of-the-art performance.

    5.Limited ablation and sensitivity analysis The paper provides some ablation on the placement of CorBN layers, but lacks deeper analysis of key design choices. For example, there is no exploration of sensitivity to the chosen metric (e. g. , Off-Log vs. alternatives), the effect of scale and bias parameters, or comparisons with simpler normalization strategies. This makes it difficult to fully understand the contribution of each component.

    6.Weak clinical relevance and application impact Although EEG is mentioned as a clinical modality, the paper focuses on standard decoding tasks (e. g. , motor imagery, SSVEP, ERN) rather than clinically meaningful outcomes. There is no demonstration of clinical feasibility, diagnostic utility, or integration into medical workflows, which limits its impact for application-oriented venues such as MICCAI.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    On the positive side, the paper addresses a clear and meaningful problem: existing normalization techniques are not directly applicable to correlation matrices due to their manifold constraints. The proposed Correlation Batch Normalization (CorBN) is well-motivated and derived in a principled way using the Lie group structure of the correlation manifold. The method is mathematically clean, includes closed-form formulations, and is supported by theoretical justification. In addition, the paper is clearly written and demonstrates consistent improvements in training stability and convergence behavior.

    However, the main limitation lies in the level of novelty. The proposed method is largely an extension of existing ideas, such as Riemannian batch normalization for SPD matrices and Lie group normalization methods, adapted to a specific manifold. While this extension is non-trivial, it remains incremental rather than fundamentally new.

    Furthermore, the empirical improvements are relatively small in magnitude, typically around 0.5–1%, and the evaluation is conducted on a limited set of standard EEG benchmarks. The lack of broader validation on larger or more clinically relevant datasets reduces the overall impact. In addition, comparisons with more recent state-of-the-art methods and deeper ablation studies are limited, making it difficult to fully assess the advantages of the proposed approach.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    Thank you to the authors for the detailed and constructive rebuttal. The response addressed several of my main concerns, particularly regarding the manifold-specific nature of CorBN, the statistical significance of the improvements, and the broader validation on larger clinical EEG datasets.

    I still view the contribution as somewhat incremental relative to prior Riemannian and Lie-group normalization approaches. However, the rebuttal clarified that CorBN is not a straightforward transfer of existing SPD-based BN methods, since the correlation manifold introduces unique geometric constraints and requires manifold-specific normalization operations under the OLM geometry.

    I also appreciate the additional evidence regarding stability, Wilcoxon significance tests, component-wise ablations, and the extended validation on TUAB with LaBraM-base. These additions strengthen the empirical support and improve confidence in the practical relevance of the method.

    Overall, while my concerns regarding the level of methodological novelty remain to some extent, I believe the rebuttal sufficiently addressed the major weaknesses raised in my review. I am therefore updating my score from Weak Reject to Weak Accept.



Review #2

  • Please describe the contribution of the paper

    The paper proposed a principled normalization framework for classification of EEG signals.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Novel Theoretical Foundation with Riemannian Geometry 2.Closed-Form Computational Efficiency 3.Applicable to different EEG datasets.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The accuracy improvement is incremental. The authors are suggested to discuss the limitations of the method, especially in which cases the method does not work.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The method has novel theoretical foundation with Riemannian geometry and closed-form computational efficiency.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors addressed the concerns.



Author Feedback

We thank the reviewers and AC for the very constructive comments.

1.@R1@AC: Novelty. CorBN is related to prior Riemannian/Lie-group normalization, but is not a direct transfer of existing BN. Correlation matrices form a strict SPD submanifold with a unit-diagonal constraint; thus Euclidean or SPD-based BN may violate valid Correlation geometry. Moreover, BN operations, including centering, scaling, biasing, and running-statistics updates, are metric- and manifold-dependent. Under OLM, they must be re-derived using the specific Lie group structure, where the logarithmic map lies in the hollow symmetric space and the exponential map requires diagonal correction to preserve unit diagonals. CorBN thus provides a missing normalization primitive for Correlation-valued neural networks. It defines OLM-consistent batch statistics, centers samples by group translation, controls dispersion in the OLM tangent space, and applies learnable bias while preserving the manifold constraint. This is important for EEG decoding, where Correlation matrices capture scale-invariant inter-channel dependencies beyond covariance-based SPD descriptors. CorBN also benefits from closed-form OLM-related operators and trivialization-based optimization [NeurIPS, 2019], making it a practical step toward deep learning on Correlation manifolds.

2.@R1@R2@AC: Accuracy. However, EEG decoding is noisy and subject-dependent, where consistent improvements are non-trivial. CorBN improves stability, reducing standard deviations from 2.78 / 3.22 / 3.40 to 1.98 / 2.82 / 3.27.Compared with SPDNetBN, CorBN achieves larger gains of +2.47%, +5.62%, and +7.25%. Subject-level paired Wilcoxon tests further support significance over CorAtt: MI (p=0.0039), SSVEP (p=0.0059), and ERN (p=0.000214).

3.@R1@AC: Dataset. We additionally evaluate CorBN on TUAB, a clinical abnormal-EEG benchmark [Frontiers in Neuroscience, 2016] with over 400K EEG samples. We also compare with LaBraM-base, a recent EEG foundation model [ICLR, 2024], which reports 0.9022±0.0009 AUROC. Under the same setting, LaBraM-base+SPDNetBN obtains 0.9034±0.0021, while LaBraM-base+CorBN achieves 0.9098±0.0011.These show CorBN remains effective on a larger clinical dataset and works as a pluggable module that can be compatible with large EEG model.

4.@R1: Ablations. We add component-wise ablations on the three used datasets. Centering alone achieves 75.01±1.68 / 68.09±2.73 / 75.57±3.63, Centering+Scaling improves to 75.12±2.49 / 68.24±3.00 / 75.70±3.54, while Full CorBN: Centering+Scaling+Bias, obtains the best accuracy. These show that each component contributes and the learnable bias helps restore a task-adaptive manifold mean after normalization.

5.@R1: Metric choice. OLM is not arbitrary. It is a permutation-invariant log-Euclidean geometry designed for Correlation matrices [SIAM J. Matrix Anal. Appl., 2024]. Its Riemannian operators naturally support CorBN computations while preserving the unit-diagonal constraint. In contrast, LSM involves the log-scaling bijection and diagonal matrix-scaling, making the BN pipeline less direct. We choose OLM for its simpler and implementation-friendly geometry, and will add detailed analysis in the revision.

6.@R1@R2: Reproducibility. Once accepted, we will release all the implementation codes and files on GitHub as soon as possible.

7.@R2: Limitations. CorBN is designed for scenarios where scale-invariant inter-channel dependency is discriminative. It may be less effective when absolute magnitude, channel variance, or covariance energy carries task-relevant information, since Correlation representations remove scale information. On the RADAR dataset, where covariance magnitude reflect target reflection intensity, motion strength, distance attenuation, and signal energy, SPDNetBN outperforms CorAtt-BN (94.85±0.99 vs 93.89±2.11). This suggests CorBN is not universally preferable and seems to be more beneficial for EEG-like signals dominated by dependency pattern.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    This paper proposes a Riemannian batch normalization on Correlation manifolds for EEG decoding. This extension is technically meaningful. However, the main limitations are the level of novelty and the empirical gains. The proposed method largely extends existing ideas, while non-trivial, its contribution appears incremental rather than fundamentally new. The empirical gains are also relatively modest, typically around 0.5~1%, and the evaluation is limited to a small set of standard EEG benchmarks. All these concerns and others raised by the reviewers should be addressed in the rebuttal.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    After considering the reviewers’ updated assessments and the authors’ rebuttal, I agree with the recommendation to accept the paper. Although the concerns regarding the degree of methodological novelty are not fully eliminated, the rebuttal sufficiently addresses the major weaknesses raised in the first round, and both reviewers have updated or maintained positive recommendations. Thus, I recommend acceptance.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Reviewer concerns seem to have been addressed adequately in the rebuttal response. Overall, the paper appears to offer a good balance between methodological novelty, appropriateness for the application, and empirical justification.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper proposes CorBN, a Riemannian batch normalization framework designed for correlation manifolds to improve EEG decoding while preserving the geometric constraints of full-rank correlation matrices. The work addresses a clear methodological gap by introducing a geometrically principled and computationally practical normalization layer that is tailored to correlation-valued representations, with a sound theoretical formulation and consistent empirical improvements across multiple EEG decoding tasks. While the contribution is somewhat incremental relative to prior Riemannian and Lie-group normalization methods and the performance gains are modest, the rebuttal sufficiently clarified the manifold-specific necessity of the approach, strengthened the empirical justification through additional analysis, and discussed limitations more clearly. Excluding the additional TUAB evaluation results introduced in the rebuttal and basing this assessment on the submitted paper together with rebuttal clarifications, I find that the main concerns are sufficiently mitigated. Overall, the work is technically solid, well motivated, and provides a useful normalization primitive for correlation-manifold learning.



back to top