- 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
- 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.
- 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
-
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.
-
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
-
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.
-
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?
-
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.
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=−ΔHt yields R=0 while simply returning Ht−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.