acceptodds
Under review as a conference paper at ICLR 2027

Counteracting Detrimental Propagation across Space and Time in EEG

Abstract

Electroencephalography (EEG) enables non-invasive, real-time recording of neural activity, with broad applications in neurological diagnosis, monitoring and rehabilitation. Yet, reliable EEG analysis is hindered by task-irrelevant signals that propagate across electrodes, accumulate over time, and become entangled with predictive cues. Existing attempts focus on the outcome level by regularizing resulting representations or structures, leaving the propagation process largely unaddressed. By the time they intervene, detrimental signals may already have spread and become entangled with predictive cues, making post-hoc removal intrinsically difficult and risking collateral damage to task-relevant neural signatures. In this paper, we propose CounterField, which recasts the mitigation of task-irrelevant signals as counteraction against their propagation from a graph stochastic partial differential equation (SPDE) perspective. CounterField structures a spectral response against detrimental propagation, enforcing coherent counteraction while preserving predictive flow. We extensively evaluate CounterField across diverse EEG tasks, where consistent gains demonstrate the effectiveness of propagation-level counteraction in learning compact yet predictive spatiotemporal dynamics.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
EEG, Spatiotemporal Graph Learning, Graph Diffusion, Compact Graph Representation Learning, Epileptic Seizure Detection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_N5gzOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes CounterField, a framework for EEG representation learning that aims to suppress task-irrelevant signals during their spatiotemporal propagation rather than refining the resulting representations afterward. The method first constructs dynamic inter-electrode graphs from frequency-wise amplitude agreement and phase compatibility. It then applies a Matérn-inspired spectral operator to consecutive-state transitions, producing a counteractive residual intended to cancel detrimental components. A log-determinant parsimony objective encourages compact representations, while a gradient-based sufficiency objective is intended to preserve task-relevant information.

The method is evaluated on four EEG applications: seizure detection and prediction on TUSZ, emotion recognition on FACED, sleep staging on ISRUC, and imagined-speech classification on BCIC2020-3. The reported results generally improve over the selected baselines. The paper also provides component ablations and analyzes the effects of the parsimony and sufficiency coefficients on predictive performance and the magnitude of counteraction.

Critical strengths
  1. The paper targets an important and practically relevant EEG problem.

EEG signals contain substantial subject-, sensor-, and task-irrelevant variation, and such variation can become mixed across channels and time. Considering when and where nuisance suppression is applied, rather than only increasing backbone capacity, is a meaningful research direction.

  1. The empirical evaluation covers several distinct EEG applications.

The paper evaluates seizure analysis, emotion recognition, sleep staging, and imagined-speech classification and reports generally favorable results across these settings. The coefficient analyses also provide useful initial evidence that the compactness and task-preservation objectives regulate different aspects of the learned correction.

Critical weaknesses
  1. The claimed SPDE and Gaussian-random-field interpretation is not consistent with the implemented formulation. The paper initially motivates CounterField using a stochastic differential equation with a stochastic term, but the core derivation subsequently uses a deterministic state equation. The implemented counteractive field is also a deterministic spectral transformation of the observed state difference, with no stochastic sampling procedure, random field realization, or probabilistic training objective. In addition, the spectral operator contains a signed learned modulator. If this modulator can take negative values, the resulting operator is not necessarily positive semidefinite and therefore cannot generally be interpreted as the covariance of a Gaussian random field. As currently presented, the method appears closer to a learned graph-spectral residual filter than to the claimed graph-SPDE random field. This issue affects both the correctness and novelty of the central theoretical interpretation.

  2. The paper does not establish the claimed distinction between propagation-level counteraction and outcome-level residual refinement. Equation (19) explicitly shows that the refined transition is algebraically equivalent to adding the counteractive correction to the current state. Moreover, the correction is computed only after the complete transition between two observed states is available. It is therefore unclear in what sense the method intervenes before detrimental information propagates. The predictive and task-irrelevant components of a transition are also latent and unidentifiable. The parsimony loss encourages contraction, while the sufficiency loss constrains only a first-order approximation to the task-loss change. These objectives do not by themselves demonstrate that the removed component corresponds to task-irrelevant propagation. The absence of parameter-matched residual and generic graph-filter controls makes it difficult to attribute the reported gains to the proposed mechanism rather than to additional capacity or generic spectral regularization.

Questions for authors
  1. Where does randomness enter the implemented CounterField model? The proposed counteractive operator appears deterministic, while Equation (19) makes the method algebraically equivalent to adding a graph-spectral residual to the current state. Please clarify what formally distinguishes CounterField from a deterministic residual graph filter. If the spectral operator is intended to represent the covariance of a Gaussian random field, please also explain how positive semidefiniteness is guaranteed when the learned spectral modulator can be signed.

  2. Are all four evaluations strictly patient- or subject-independent? Please specify the exact unit used to construct the training, validation, and test splits and confirm that no clips or windows from the same patient or subject can occur in multiple splits. Please also confirm that normalization, automatically calibrated quantities, and all other dataset-level statistics are computed using training data only. For Figure 2, is the best epoch selected exclusively using validation performance, without selecting retrospectively from the test trajectory?

  3. Can the authors provide a parameter-matched residual or generic Chebyshev graph-filter baseline using the same backbone, preprocessing, training budget, and model-selection protocol? Please also clarify which results in Tables 1 through 4 were reproduced in the authors' pipeline and which were obtained from prior publications. These controls are important for determining whether the gains arise specifically from transition-based counteraction rather than additional capacity, backbone differences, or generic spectral filtering.

Reviewer_Few4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that task-irrelevant signals in EEG propagate across the electrode graph and accumulate over time, so they should be suppressed during propagation rather than compressed in the final representation. The method first uses STFT magnitude and phase to construct a dynamic electrode graph for each snapshot. From a graph SPDE perspective, it then adds a counteractive field to the transition between adjacent states, (ΔHt\Delta H_t). The field is parameterized by a Matérn spectral function and approximated with Chebyshev polynomials. The method also uses a log-det-based parsimony loss and a task-gradient-based sufficiency loss. The authors add it as a plug-in to EvoBrain and CBraMod, report improvements on four tasks using TUSZ, FACED, ISRUC, and BCIC2020-3, and compare it with outcome-level methods such as GIB and graph condensation.

Critical strengths

Intervening in state transitions rather than compressing the final representation is an intuitive approach. As a plug-in, the method is evaluated with both a Mamba backbone (EvoBrain) and a Transformer backbone (CBraMod) across four types of EEG tasks. Table 4 compares the method with GIB, graph condensation, and their dynamic variants under the same backbone. This is the comparison most directly related to the paper’s central claim.

Critical weaknesses

There is a substantial gap between the paper’s SPDE framing and what the method actually does. The implementation has no stochastic term. The final operation is simply (H~t=Ht+∑kckTk(L~ttrans)ΔHt\tilde H_t = H_t + \sum_k c_k T_k(\tilde L_t^{\mathrm{trans}})\Delta H_t). A freely learnable (mϕ(λ)m_\phi(\lambda)) is added to the Matérn term and largely determines the spectral shape, so the method is essentially a learnable Chebyshev graph filter applied to temporal differences. The paper also provides no experimental evidence that the counteracted signal is task-irrelevant; Figure 3 only shows that the regularizers have an effect. The ablations consist of just two coarse settings, SpecSL-only and CCL-only, with no comparison against an ordinary residual graph filter of comparable capacity.

Questions for authors
  1. Why do the numbers in Tables 2 and 4 differ? Were the baselines rerun with the same seeds, or were their results taken directly from the original papers?

  2. CBraMod has no inherent graph structure. At which layer is CounterField inserted, and where does HtH_t come from? What is the form of Φθ\Phi_\theta, and how are ηm\eta_m, κm\kappa_m, μm\mu_m, γ\gamma, τ\tau, and ωb\omega_b set?

  3. Could the authors add comparisons that remove the Matérn term and retain only learnable Chebyshev coefficients; apply the filter to HtH_t instead of ΔHt\Delta H_t; replace LtransL_{\mathrm{trans}} with a static Laplacian; and add an ordinary layer with a comparable parameter count to the baseline? These would help identify which component produces the improvement.

  4. Is there more direct evidence that the counteracted signal is task-irrelevant? For example, the authors could inject EOG/EMG artifacts or bad channels into TUSZ or FACED and test whether CounterField is substantially more robust.

  5. Each snapshot requires N2MN^2M pairwise spectral evidence computations and LtransL_{\mathrm{trans}}, while LsufL_{\mathrm{suf}} also requires task gradients. What are the additional training and inference costs in time and memory?

  6. Could the authors provide complete ablations on all four datasets?

  7. Given that the model has no stochastic term, why is it described as an SPDE? What does this framework contribute to the design that an ordinary graph filter could not provide?

Reviewer_Tw1JOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CounterField, which recasts the mitigation of task-irrelevant signals as counteraction against their propagation from a graph stochastic partialdifferential equation (SPDE) perspective. CounterField structures a spectral re-sponse against detrimental propagation, enforcing coherent counteraction whilepreserving predictive flow. The paper extensively evaluates CounterField across diverseEEG tasks, where consistent gains demonstrate the effectiveness of propagation-level counteraction in learning compact yet predictive spatiotemporal dynamics.

Critical strengths
  1. The paper presents an interesting perspective on mitigating task-irrelevant information in EEG signals by focusing on the propagation process rather than solely on the compression of final representations. By integrating a graph-SPDE-inspired formulation with a Matérn spectral operator, the proposed framework introduces a structured counteractive mechanism. The parsimony and predictive sufficiency objectives further aim to suppress redundant information while preserving task-relevant features. Overall, the methodological design is coherent and reasonably motivated.
  2. The paper conducts experiments across diverse EEG tasks, including seizure analysis, emotion recognition, sleep staging, and imagined speech classification, and reports generally consistent improvements over competitive baselines. Furthermore, the ablation studies, convergence analysis, and visualizations under different regularization settings provide additional empirical support for the practical effectiveness of the proposed components.
Critical weaknesses

1.Insufficient direct validation of detrimental propagation suppression. The central claim of the paper is that CounterField actively suppresses task-irrelevant information during spatiotemporal propagation. However, the existing experiments primarily provide indirect evidence through downstream performance improvements, counteraction magnitude, and predictive sufficiency measurements. There is limited direct evidence demonstrating that detrimental propagation is actually suppressed. Consequently, it remains unclear whether the proposed method specifically removes task-irrelevant information or merely improves predictive performance through general feature transformations or regularization. Controlled perturbation experiments, together with quantitative analyses or visualizations of nuisance propagation and attenuation, would help substantiate this central claim. 2.Unclear distinction between propagation-level counteraction and conventional residual refinement. The paper emphasizes intervention during information propagation rather than post-hoc representation refinement as a key contribution. However, Eq. (19) shows that the proposed propagation-level intervention can be algebraically expressed as a residual correction to the current state. Although this equivalence does not invalidate the method, the substantive distinction from conventional graph-spectral filtering or residual refinement remains insufficiently established. Moreover, the lack of controlled comparisons with such alternatives under comparable backbones, model capacities, and computational budgets makes it difficult to isolate the contribution of the proposed propagation-level formulation. 3.Limited controlled evidence for cross-backbone applicability. The paper reports consistent improvements across different EEG tasks and backbone architectures. However, Mamba is primarily employed for seizure analysis, whereas Transformer is used for the remaining tasks, making backbone architecture and task configuration confounded in the experimental design. Although the reported results demonstrate that CounterField can be applied to different tasks and architectures, they do not fully disentangle the effects of backbone choice and task characteristics on the observed improvements. Therefore, the current evidence is insufficient to rigorously establish whether the performance gains are consistent across backbones under the same task setting. Controlled comparisons applying CounterField to different backbones on the same dataset would strengthen the claimed cross-architecture applicability.

Questions for authors

Q1. Direct validation of detrimental propagation suppression.The central claim of CounterField is that it actively suppresses task-irrelevant information during spatiotemporal propagation. However, the current experiments mainly provide indirect evidence through downstream performance improvements and counteraction-related measurements. Could the authors provide more direct evidence that the learned counteractive field specifically suppresses detrimental propagation? For example, can controlled perturbation experiments or visualizations demonstrate how task-irrelevant signals propagate and are attenuated before and after counteraction? Q2. Distinction from conventional residual refinement.According to Eq. (19), the proposed propagation-level counteraction can be equivalently expressed as a residual correction to the current state. What distinguishes CounterField from conventional graph-spectral filtering or residual refinement, both theoretically and empirically? Could the authors clarify the unique contribution of the propagation-level formulation and provide controlled comparisons with simpler alternatives under comparable model capacities and computational budgets? Q3. Cross-backbone applicability.The paper reports consistent improvements across different EEG tasks and backbone architectures. However, Mamba and Transformer are evaluated on different tasks, making it difficult to disentangle the effects of backbone architecture and task characteristics. Could the authors clarify whether CounterField yields consistent improvements when different backbones are evaluated on the same dataset? If such experiments have been conducted, providing the corresponding results would help substantiate the claimed cross-backbone applicability.

Comments on the paper

1 comment

Thanks for the response and additional experiments! The matched controls and cross-backbone results strengthen the empirical evidence. However, the distinction from regularized transition filtering remains insufficiently established. In particular, setting Ct=−ΔHtC_t=-\Delta H_t yields R=0R=0 while simply returning Ht−1H_{t-1}, so this metric alone does not demonstrate selective nuisance removal.

Please include the “w/o Matérn” variant, retaining both regularizers, in the noise analysis. Please also clarify whether R2’s clean/noisy results use identical checkpoints and reconcile the ISRUC standard deviations in R7 with Table 2.

I am maintaining my current rating of weak rejection, but I am open to raising it if the concerns above are convincingly addressed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.