acceptodds
Withdrawn from ICLR 2027

FROM PHYSIOLOGICAL COMPOSITION TO AFFECTIVE GEOMETRY IN EEG REPRESENTATIONS

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Emotion is associated with joint activity across multiple brain regions. However, the regional and spectral composition of emotion-related activity remains difficult to characterize explicitly in high-dimensional EEG representations. In this work, we propose a physiological compositional framework that expresses emotion-related EEG activity through a shared vocabulary of region–frequency motifs. For each EEG instance, we estimate motif activations from EEG foundation-model representations, guided by each motif’s predefined scalp region and spectral activity within its frequency band. Competition among motifs reduces redundant evidence assignment, and their activations together form an instance-specific physiological composition. The resulting representation provides a shared physiological space for comparing affective geometries across foundation models. Across all nine dataset–model combinations, physiologically structured motifs outperform randomized mappings, supporting the presence and predictive relevance of physiological organization in the learned composition. And two independently pretrained foundation models also recover highly similar affective geometry, with similarities of 0.976 in emotion distances and 0.935 in centroid angles. Together, these results establish physiological composition as a common, interpretable space for comparing EEG foundation models not only by decoding performance, but also by how they organize affective structure.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Affective Computing, Representation Learning, EEG Emotion Recognition, EEG foundation models, Physiological Composition, Affective Geometry

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_4bePOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a physiological compositional framework for EEG signal analysis. It estimates motif activations from EEG foundation-model representations for each instance, guided by each motif’s predefined scalp region and spectral activity within its frequency band. The resulting representation is used for emotion prediction evaluated on SEED, SEED-IV, and a five-class in-house dataset.

Critical strengths
  1. It proposes a shared 8-region × 5-band vocabulary that expresses heterogeneous EEG foundation models in a common physiological coordinate system.
Critical weaknesses
  1. The claimed contributions are difficult to verify and of limited value to the affective computing community, as all findings are based on only three frozen foundation models.
  2. This manuscript evaluates only three foundation models with relatively low accuracies, without comparing to any other state-of-the-art emotion-recognition models.
  3. Little information is provided about the in-house dataset.
Questions for authors
  1. The work reads as an analysis of three specific foundation models rather than a contribution to emotion research. The authors should clarify what generalizable insight about emotion representation is being offered, and demonstrate that it holds beyond this model set and dataset.
  2. More state-of-the-art emotion-recognition models should be compared using the proposed EEG representation to evaluate its effectiveness comprehensively.
  3. A link of the in-house dataset should be provided in this submission, or at minimum detailed subject, trial, preprocessing, and fold-split information should be provided.
Reviewer_ukcaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper maps frozen EEG foundation-model features into 40 predefined scalp-region and frequency-band motifs. A supervised downstream pathway combines learned queries, spatial and raw band-power guidance, and competition among motifs. Motif activations contribute additively to emotion logits alongside a base predictor. Across three datasets and three backbones, the authors evaluate predictive utility, randomized mappings, and motif dependence. They then compare emotion-centroid geometries, finding strong LaBraM-REVE correspondence on a five-class in-house dataset.

Critical strengths
  1. A shared vocabulary offers a useful way to compare heterogeneous EEG representations beyond classification accuracy.
  2. The analysis includes several controls that go beyond visual inspection: subject-conditioned and exact-count null models, emotion-identity permutation tests with FDR correction, and a backbone that does not show the same geometric agreement.
Critical weaknesses
  1. The physiological meaning of individual motifs has not been validated.​ Equation 3 gives each motif soft spatial and spectral biases while retaining freely learned queries over contextual backbone features. Those features can contain information from outside the nominal region or frequency band. Assigning a region-band name to a coordinate therefore does not establish what its activation measures. The randomized-mapping control tests the predictive value of the imposed organization, but does not show that individual motifs faithfully represent their assigned physiological components.

  2. Direct spectral inputs confound the interpretation of the backbone representations. ​The motif pathway receives raw EEG band-power information in both the attention scores and the activation function, while the base predictor does not receive that explicit input. Improvements over the base can therefore arise from access to additional spectral features. They do not isolate physiological information recovered from the frozen backbone.

  3. The geometric analysis does not establish the added value of a shared motif vocabulary.​ The reported cross-model comparison uses distances and angles between emotion centroids. These quantities are invariant to orthogonal changes of feature basis, so their comparison does not require semantically aligned coordinates across models. The same RSA analysis can be performed in each backbone's native representation or in the base-branch features. Without these controls, the paper has not shown whether the motif mapping reveals additional geometric correspondence.

  4. Shared downstream supervision remains an unresolved explanation for the cross-model agreement.​ LaBraM and REVE are mapped through trainable pathways using the same emotion labels, motif definitions, and spectral inputs. Their agreement may therefore reflect structure learned during downstream classification. CBraMod's disagreement shows that convergence is not inevitable, but does not determine how much of the LaBraM-REVE correspondence comes from pretraining. The raw band-power geometry is also an incomplete control.​

  5. The stability and uncertainty of the headline geometric result are insufficiently characterized.​ The main correspondence comes from five emotion centroids on one in-house dataset, yielding ten pairwise distances. The exact emotion-identity permutation tests and FDR correction are useful, but do not quantify variability across participants, stimuli, or model fits. Fold-wise stability is examined for CBraMod, while comparable results for LaBraM and REVE are missing.

  6. The dependence and intervention analyses provide limited independent validation of physiological composition.​ The total-correlation nulls address marginal activation rates, subject tendencies, and active-motif counts. They do not exclude dependence induced by shared emotion or stimulus inputs, common spectral fluctuations, or supervised training. Thus, significant dependence does not by itself identify a physiological compositional mechanism. The intervention analysis has a separate limitation: with additive logits, removing a motif measures its class-margin contribution, and the necessity objective explicitly encourages that contribution to be positive.

Questions for authors
  1. What evidence shows that each motif responds selectively to its named region and frequency band?

  2. How do prediction and geometry change when direct band-power inputs are removed? What are the corresponding LaBraM-REVE RSA results in native backbone and base-branch representations? Does the motif mapping reveal correspondence that these representations do not already exhibit?

  3. Can matched supervised spectral or untrained-backbone controls reproduce the cross-model agreement?

Reviewer_e9beOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a physiological motif framework for analyzing EEG foundation-model representations. Forty motifs are defined by combining eight scalp regions with five frequency bands. A trainable pathway uses structured queries, competitive evidence assignment, and gating to combine frozen foundation-model features with spectral information computed directly from EEG. The resulting motif activations contribute to emotion prediction and provide a representation in which emotion-centroid geometries are compared across models. Experiments cover three datasets and three foundation models, reporting predictive advantages over randomized mappings and strong geometric correspondence between LaBraM and REVE on an in-house five-class dataset.

Critical strengths

1.The research question is meaningful. The paper seeks to understand EEG foundation models through physiological interpretation and representation structure, providing a useful complement to evaluations based primarily on classification accuracy. 2.The empirical evaluation has reasonable breadth. The study includes multiple datasets and backbones, randomized-mapping controls, architectural ablations, and statistical null models. It also reports disagreement between models, providing an informative initial assessment of the proposed framework.

Critical weaknesses

1.The central physiological interpretation is not sufficiently validated.

The paper treats each motif as a coordinate with a fixed region–frequency meaning. However, the spatial and spectral terms in Equation (3) are soft biases, the queries contain free parameters, and the final activations are produced by trainable functions supervised through emotion labels. A coordinate can therefore encode discriminative information that is not specific to its designated region or frequency band.

The randomized-mapping comparison in Table 1 supports the predictive usefulness of the physiological prior, while the attention-redundancy results in Table 5 support differentiated evidence assignment. Neither directly validates the physiological semantics of individual coordinates. Similarly, the interventions in Equation (6) precisely measure a scalar’s contribution to the prediction, but do not independently establish its physiological meaning. This gap affects the paper’s central contribution: whether the learned representation provides a reliably interpretable physiological coordinate system.

2.The source of the cross-model geometric correspondence remains insufficiently controlled.

Both LaBraM and REVE motif representations are trained using the same emotion labels and downstream objectives, and both receive spectral information computed directly from the EEG. The high correlations in Table 3 may therefore reflect the combined influence of shared supervision, shared spectral information, and the downstream mapping. The geometry computed directly from raw band-power features is not a supervision-matched control. The disagreement of CBraMod establishes that the shared procedure does not inevitably produce agreement, but does not isolate the factors responsible for the LaBraM–REVE result.

Moreover, pairwise distances and centroid angles are invariant to coordinate permutations. These metrics can consequently agree even when identically named motifs do not have corresponding physiological meanings across models. The current evidence does not sufficiently establish the specific value of the proposed physiological coordinates over an equally supervised generic representation or a learned spectral representation for cross-model comparison.

3.The main positive findings lack adequate stability and uncertainty analyses.

The headline geometric result is based on five emotion categories in the in-house dataset, yielding ten pairwise distances. The exact emotion-identity permutation tests and multiple-comparison correction support the statistical significance of the correspondence between the estimated geometries. However, they do not characterize uncertainty from participant and stimulus sampling or from downstream training randomness. Figures 4 and Table 4 examine fold-wise variability only for CBraMod; equivalent analyses are not reported for the central positive LaBraM–REVE finding.

In addition, the accuracy advantages of structured over randomized motifs in Table 1 range from 0.4 to 3.6 percentage points, with several small differences. Variability across random mappings and training seeds is not reported. It is therefore difficult to assess the reproducibility of both the physiological-mapping advantage and the principal geometric finding.

4.Essential implementation and evaluation details are missing.

Several choices that directly determine the results are not sufficiently specified, including the motif activation function, the precise intervention losses and their weights, the recurrent-motif selection rule, and the construction of randomized mappings. EEG preprocessing, feature-window settings, and the exact evaluation splits for SEED and SEED-IV also require further detail.

These omissions prevent a clear assessment of whether the randomized control changes only the intended physiological correspondence while preserving available information and structural complexity. They also leave unclear how activations from independently trained outer-fold motif heads are normalized and combined into a common geometry. Because the main cross-model finding additionally relies on an incompletely described in-house dataset, independent verification is difficult.

Questions for authors

1.Physiological semantics: What direct evidence, beyond classification accuracy and attention redundancy, demonstrates stable region–frequency specificity of individual motifs? Could controlled region–band perturbations or independent physiological measurements validate the intended correspondence and establish that identically named motifs have consistent meanings across models?

2.Attribution of geometric agreement: Under matched capacity and supervision, how much geometric correspondence is obtained from a generic low-dimensional representation or a learned band-power-only representation? Does the main finding persist when direct spectral inputs from the raw EEG are removed?

3.Stability: Please report LaBraM–REVE geometric correspondence separately across outer folds and training seeds, together with uncertainty accounting for participants and shared video stimuli. Is the predictive advantage of structured mappings also consistent across multiple fixed random mappings and matched training seeds?

4.Evaluation protocol: How are randomized mappings constructed, and do they preserve region sizes, overlap, and frequency-band coverage? How are recurrent motifs selected? How are motif activations from separately trained outer folds normalized and combined to ensure that the resulting geometry has comparable coordinate meanings?

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

Related papers

Discussion (0)

Sign in to comment.