acceptodds
Under review as a conference paper at ICLR 2027

SPDAlign: Interpretable Riemannian Alignment for EEG Forward Modeling Shifts

Abstract

Electroencephalography (EEG) based brain-computer interfaces enable direct brain-to-device communication for applications such as rehabilitation and communication. However, their practical utility is often limited as the non-stationary nature of the EEG data introduces distribution shifts across domains (e.g., sessions and subjects). Adapting machine learning models to be invariant to these shifts in an unsupervised way, without using costly labeled calibration data, would drastically improve the utility of EEG data. In this work, we use a classic generative model of EEG to study distribution shifts introduced by the domain-specific forward process, which is associated with factors such as head geometry. We theoretically show that such distribution shifts can be recovered solely through linear transformations on the Symmetric Positive Definite manifold. Building on this insight, we propose SPDAlign, an interpretable framework for promoting domain-invariant EEG learning. SPDAlign first aligns the domain-specific means and corrects global rotations across domains using a recent optimal transport technique called Wasserstein Procrustes. We systematically study the proposed approach through simulations and demonstrate its competitive performance on extensive public EEG datasets. Additionally, SPDAlign is a globally linear framework and is intrinsically interpretable, so that the framework can identify frequency ranges of interest, determine the spatial patterns reflecting source–sensor relationships, and address cross-subject variability.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Electroencephalography, geometric deep learning, unsupervised domain adaptation, brain-computer interfaces

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_b9MrOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This submission focuses on challenges around brain-computer interfaces (BCIs) and presents SPDAlign, an interpretable end-to-end Riemannian learning framework designed to mitigate distribution shifts in electroencephalography (EEG) data, which frequently experience distribution shifts across sessions and subjects due to variations in domain-specific forward processes, such as head geometry. The submission theoretically demonstrates that these shifts can be recovered entirely through linear transformations on the Symmetric Positive Definite (SPD) manifold, specifically involving mean matching followed by a global rotation transformation. To achieve unsupervised multi-source, multi-target domain adaptation, SPDAlign first aligns domain-specific means and then corrects global rotations across domains by estimating an orthogonal transformation via an optimal transport technique called Wasserstein Procrustes. The proposed framework is validated using simulations and extensive cross-subject decoding experiments across four public benchmark EEG datasets: motor imagery, workload estimation, mental stress, and driving fatigue. Because the framework maintains global linearity, it is intrinsically interpretable; classifier weights can be mapped directly back to sensor space to identify relevant frequency ranges (such as the theta band) and specific brain regions reflecting source-sensor relationships.

Critical strengths
  1. The paper provides a clear mathematical bridge between the physical linear generative process of EEG (forward modeling) and geometric operations on the Symmetric Positive Definite (SPD) manifold. It proves that distribution shifts caused by subject-specific head geometry can be recovered via mean alignment and global rotations. Unlike existing Riemannian rotation alignment methods that rely on target domain labels, SPDAlign is capable of estimating global rotations in an entirely unsupervised manner using Wasserstein Procrustes.

2) Because the framework is globally linear, it avoids the "black box" nature of deep learning. Classifier weights can be mapped directly back to sensor space, successfully identifying neurophysiologically meaningful markers (e.g., frontal/parietal theta activity for mental workload). It naturally scales to multi-source and multi-target unsupervised domain adaptation (UDA), which directly addresses real-world BCI deployment where data from multiple historical subjects must be adapted to a new user seamlessly.

  1. The framework demonstrates superior generalizability by achieving state-of-the-art cross-subject decoding results on 3 out of 4 extensive public benchmark datasets (Workload, Stress, and Fatigue), outperforming competitive deep learning and foundation models.
Critical weaknesses
  1. The core computational limitations of SPDAlign stem from its reliance on Riemannian geometry and iterative optimal transport. Because the framework operates directly on the Symmetric Positive Definite (SPD) manifold, it may scale poorly with both the dimensionality of the feature space and the number of target samples. Due to this reason, SPDAlign may limit its applicability to high-dimensional SPD features.

  2. There remains a pronounced performance gap between the unsupervised SPDAlign and its "Oracle" counterpart. This discrepancy stems from ambiguous transport plans when class-conditional distributions overlap, leading to suboptimal rotation estimation. In addition, simulation results highlight that the unsupervised rotation estimation degrades significantly under extreme rotation shifts or weak class separability, making it less robust in highly ambiguous noise regimes.

  3. While the authors suggest that the framework can be extended to online UDA via running statistics, the study is entirely restricted to offline dataset evaluations, leaving its performance in interactive, real-time BCI loops unverified.

  4. Reliance on Linear Assumptions: The framework's mathematical proofs rely on standard simplifying assumptions of a linear, stationary generative EEG model with uncorrelated sources. This may not completely capture highly non-linear brain dynamics, sensor artifacts, or non-stationary noise drift encountered in field settings.

  5. The literature is a bit old and most recent research efforts should be investigated and included. Please also check the references, which are not presented in a consistent way.

Questions for authors

Please check the Critical Weaknesses section and clarify.

Reviewer_LdacOpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In their work “SPDAlign - Interpretable Riemannian Alignment for EEG Forward Modeling Shifts”, the authors address the important problem of distribution shifts of EEG data, which currently pose a major and so far only incompletely solved problem in the field for EEG-based Brain-Computer-Interfacing (BCI) research and application. Such shifts may occur for example between sessions of the same subjects, or between different subjects, for a broad number of reasons. SPDAlign specifically addresses unsupervised domain adaptation in the Riemannian EEG decoding context. The presented work has a theoretical, methodological, simulation-based, and an empirical part. Theoretically, the authors extend an existing theoretical framework by Li and colleagues (SPDIM) and show that under certain assumptions, forward-model-related distribution shifts can be undone by linear transformations on the SPD manifold. In the methodological part, they propose SPDAlign - combining Riemannian Alignment based on both recentering and rotation as the linear transformations following the theory, and using Wasserstein Procrustes for target alignment, together with a specific training-finetuning and decoding pipeline. In the simulation part, several aspects of the proposed methods are studied, in particular the usefulness of adding rotations on top of just recentering. For that, the simulations examine the benefits and limitations of the proposed alignment under different rotation and data conditions. Finally, in the empirical part, SPDAlign is applied to four public BCI datasets, showing competitive performance; the main comparisons against competing methods are however not accompanied by reported significance tests. Finally, examples for the interpretability of the SPDAlign approach are exemplarily provided for a workload decoding task, showing spectral profiles and scalp topographic patterns broadly in line with the literature. In summary the paper makes specific, valuable contributions to various aspects of Riemannian EEG alignment, and thus presents a potentially useful extension and integration of existing Riemannian EEG alignment methods, supported by simulations and empirical evaluation. However, most of its individual architectural and interpretability components are inherited from prior work. Also, concerns regarding the scope of the theoretical claims and strength of the evaluation remain to be addressed.

Critical strengths

A first critical strength lies in the practical value of combining unsupervised target alignment with an interpretable EEG decoding architecture, in combination with the reported results across four datasets / relevant application domains covering motor imagery, workload, stress, and fatigue, showing competitive performance without requiring target labels, while the underlying architecture offers interpretability wrt. to the underlying spectral and spatial patterns. This supports the potential usefulness and practical relevance of the proposed SPDAlign approach, even though its individual components are largely based on prior work.

A second major strength I see lies in the combination of theory, simulation, method development, and empirical evaluation - which is quite rare in the EEG-BCI literature, and which, in the best case, has the potential for promoting the field mode fundamentally. This strength, however, at least in the current state of the work, is still underdeveloped, as also discussed below.

Critical weaknesses

A major critical weakness I would see lies in the relatively incremental nature of the individual contributions, in combination with the fact that the connections between the different novel parts are less tightly integrated than the presentation in the abstract / introduction suggests. Recentring and rotational alignment, Wasserstein–Procrustes estimation, and the underlying interpretable decoding architecture taken individually are all based on prior work.

The theoretical part incrementally builds on the existing framework by Li and colleagues (SPDIM). It provides conditions under which linear transformations can undo forward-model-related distribution shifts, but the corresponding class of Riemannian EEG alignment methods already exists. This theoretical extension thus appears to provide a theoretical framework for existing alignment methods, rather than inspiring a new framework, as the writing of the manuscripts suggests.

Specifically, the theoretical analysis motivates the use of rotation in addition to recentering, but does not imply the actual novel methodological contributions in SPDAlign, such as integrating Wasserstein–Procrustes into existing Riemannian EEG alignment. The simulations, likewise, primarily examine the usefulness of adding rotation to recentering, but do not establish the specific advantage of the chosen shift estimation procedure over existing rotational alignment approaches. Thus, the connection between the theoretical result, the proposed estimation procedure, and the reported empirical benefits remains in my view insufficiently demonstrated.

Therefore, while the overall presentation positions the work as a novel theory-based alignment framework, the actual novel components appear more as valuable, but relatively independent and incremental additions to different aspects of established Riemannian EEG alignment theory and methodology.

Questions for authors

The proposed alignment procedure naturally applies to shifts due to changes in electrode positions between sessions, or to changes in head geometry in different subjects, but would eg fail if sources become electrically invisible due to cancellation in a different forward model geometry? Could the authors discuss the limitations due to the assumptions more explicitly, and thus more clarify the scope of their theoretical claims?

This would be especially interesting for changes in mental / cognitive state - the authors write: "This is a case of latent model shifting, where variations in the forward modeling process (head geometry, sensor placement, and cognitive states) induce distribution shifts in the observed sensor data (Mellot et al., 2023)." - how exactly could SPDAlign accommodate cognitive state effects?

"Changes in neural dynamics, such as impedance drift, may also occur but are typically modeled as noise (Akalin Acar & Makeig, 2013)." - true changes in neural dynamics are an important factor for distribution shifts in the BCI context, for example due to learning and adaptation in the course of BCI usage. Such changes can be systematic; I don't see in which sense these changes should be modeled as noise, or that this is common practice, or that the provided citation would substantially support this view - could the authors please clarify?

What was the optimization budget of SPDAlign? Was SPDAlign developed on the same 4 tasks which are reported in the empirical evaluation, if yes, on which data split was optimization performed and how was overfitting prevented, if no, on which other data / problems?

The discussion claims "achieving SotA results on three of four benchmark datasets", while the abstract only refers to "competitive performance"; for the underlying results there are no significance test reported - do I see this correctly that SPDAlign does not significantly outperform any of the strongest baselines in any task?

For the ablation statistic, which were the exact sets over which the permutation tests for each scenario were computed, and how was independence of the observations in the sets ensured? In particular, the reported numbers seem to imply different set sizes for the different cases. Please clarify! Were ablations only significant when pooled over the four tasks, or also within individual tasks? Across the ablations, was the training budget kept constant? For example, for ablating both moment and rotation, was the overall training budget shorter (omitting the fine-tuning phase)? Were all ablation versions fully converged after the number of training steps used, could you please show training curves? In case training is not fully converged, alignment could give an initial performance gain, which might vanish with suitable longer training - can you rule this possibility out?

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

This paper presents SPDAlign, an interpretable framework to deal with domain-invariant EEG learning. In detail, it contains a feature extractor to map EEG signals into SPD manifold, a SPDAlign module to addresses shifts from aligning means and rotations, a classifier to output predictions.

Critical strengths

1.The authors connect the cross-subject distribution shift of EEG with factors like head geometry and electrode placement, offering the alignment meaningful interpretations. The motivation is good. 2.The experiments cover numerous scenarios, including motor imagery, workload estimation, mental stress, and driving fatigue. Moreover, the performance of the proposed method seems good.

Critical weaknesses

1.The statement that “robust latent representations can be recovered solely through linear transformations”is too strong. The authors only prove the recovery under idealized conditions, where latent Frechet means are assumed to be normalized to the identity and the exact rotation is available. However, SPDAlign relies on finite-sample mean estimates and Wasserstein Procrustes. In this case, the accurate recovery of the required rotation is not established. 2.The authors mention other unsupervised domain adaption methods in the introduction section. However, limitations of these methods are not sufficiently talked to introduce SPDAlign. Moreover, the statement of “lack theoretical foundations” is very strong, methods such as Rodrigues et al. (2019) are grounded in Riemannian geometry. 3.The paper lacks evaluation when the assumptions of invertible domain-specific forward matrix, uncorrelated latent sources, and invariant conditional source-to-label relationship are violated. 4.The same category label does not mean that two trials have one-to-one corresponding latent source activities. The paper only explains category matching and downsampling, but ignores the intra-class pairing rules. Moreover, the sensitivity of the results to sample pairing and reference-domain selection is not reported. 5.Equation (12) employs matrix conjugate transformation, while Equation (6) and Algorithm 2 describe the ordinary orthogonal mapping on the point set. It is necessary to clarify whether the actual rotation acts on the matrix or point set, and prove that the implementation is consistent with the theoretical constraints.

Questions for authors

1.The authors only prove the recovery under idealized conditions, what would the recovery be like when the condition is not that ideal? 2.Please clarify the limitations of other methods that motivates the proposed method. 3.How would the performance of SPDAlign be like when the assumptions of invertible domain-specific forward matrix, uncorrelated latent sources, and invariant conditional source-to-label relationship are violated? 4.Please clarify the intra-class pairing rules, and report the sensitivity of the results to sample pairing and reference-domain selection if possible. 5.Does the actual rotation act on the matrix or point set? Is the implementation consistent with the theoretical constraints?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.