acceptodds
Under review as a conference paper at ICLR 2027

NeuroContext-TTA: Source-Anchored Context Calibration for Cross-Subject EEG Emotion Recognition

Abstract

Cross-subject EEG emotion recognition must accommodate changes in signal statistics without labels from the new subject. Test-time adaptation can exploit recent observations, but a short, correlated history may not represent every emotion class. Building target class memories from its predictions can therefore entangle subject variation with incomplete class coverage. We propose NeuroContext-TTA, a forward-only adaptation framework that separates source class structure from target context. It translates all source prototypes by a shared context-derived offset, retaining a reference for every class without assigning target pseudo-labels. Retrieval combines similarities in the EEG embedding and encoded log-covariance spaces, and a query-level reliability gate controls residual feature and logit corrections. The calibration modules are trained on source subjects and remain frozen at deployment; only the context statistics and translated references change. Across SEED, SEED-IV, FACED-3 and FACED-9, calibration of mdJPT features improves macro-F1 by 1.35, 0.54, 5.09 and 0.52 percentage points, respectively. Evaluation with five EEG backbones further yields accuracy gains in 19 of 20 backbone–task pairs. These results support source-anchored calibration as a practical way to use unlabeled context without online parameter updates.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG emotion recognition, test-time adaptation, cross-subject learning, prototype calibration, context-conditioned inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_A8oiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a framework called NeuroContext-TTA for forward-only calibration in cross-subject EEG affective recognition. The method leverages unlabeled target-domain contextual statistics to transform source-domain class prototypes, performs dual-space prototype retrieval, and applies reliability-gated feature and logit calibration. The calibration modules are trained on labeled source-domain subject data and remain frozen at deployment. Experiments on the SEED, SEED-IV, FACED-3, and FACED-9 datasets with five EEG backbones show that the method outperforms baselines with frozen backbones under multiple settings. This paper addresses a practically meaningful problem and proposes a targeted framework. However, I have major concerns regarding the novelty of the method, the validity of the contextual calibration assumption, and the sufficiency of the experimental evidence.

Critical strengths
  1. This paper raises a practical deployment problem: how to exploit unlabeled target-domain observations without online parameter updates or training with target pseudo-labels.
  2. The source-anchored prototype design avoids constructing incomplete target class memories from potentially unreliable pseudo-labels.
Critical weaknesses
  1. The methodological novelty is limited. The proposed framework combines shared prototype transformation, dual-space retrieval, and reliability-based gated residual correction, and its technical contribution can be summarized as a combination of three existing ideas. Moreover, the paper does not convincingly justify why this particular combination is necessary, or why it is superior to alternative solutions (e.g., improving representation quality and generalization).
  2. The methodology may have a major flaw. The core of the proposed method is to use context statistics from the unlabeled target domain to transform source-domain class prototypes. A key assumption underlying the effectiveness of this shifting operation is that the distribution shift between the target context mean and the source-domain class prototype mean is stable. However, the target context mean may also be affected by the proportion of emotion categories, stimulus content, and noise, which may lead to substantial differences between the source and target distributions and thus limit the effectiveness of the method.
Questions for authors
  1. Regarding Weakness 2, how does the method perform when the proportion of emotion categories changes but the distribution shift of the target subject remains unchanged?
  2. The advantage of the proposed method is that, during deployment, the backbone model does not need to be trained on the target context. However, the proposed method requires additionally training an MLP adapter for calibration. What is the advantage of doing so compared with directly optimizing the backbone model?
  3. Why is λ\lambda in Equation (1) fixed at 0.35? What is the basis for choosing this value? Similarly, why is the parameter τ\tau in Equation (5) set to 0.1?
  4. The ablation results in Table 3 show that removing different components leads to only small performance drops. Are these results statistically significant?
Reviewer_vLMQOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
1(−1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents NeuroContext-TTA, a forward-only test-time adaptation framework for cross-subject EEG emotion recognition. The proposed method leverages unlabeled target context to construct context-conditioned references while preserving source class prototypes, facilitating dual-space retrieval and reliability-guided prediction calibration. All model parameters remain frozen during deployment. Experiments demonstrate performance improvements over the corresponding unadapted models, supported by additional ablation studies and context sensitivity analyses.

Critical strengths
  1. This paper introduces a forward-only calibration framework that improves cross-subject EEG emotion recognition performance without requiring online parameter updates or target pseudo-label training, offering a computationally efficient approach to test-time adaptation.
  2. The experimental evaluation spans multiple EEG emotion recognition benchmarks and backbone architectures. The inclusion of component-wise ablation studies, context sensitivity analyses, and sample-level prediction analyses provides a relatively comprehensive empirical assessment of the proposed framework.
Critical weaknesses
  1. The effectiveness and necessity of target-specific context calibration are insufficiently justified. Equation (1) preserves the relative geometry of source class prototypes under a shared translation but provides no theoretical guarantee of effective cross-subject distribution alignment. Moreover, the context-source experiments in Appendix I show that source-derived or zero contexts can achieve performance comparable to actual target contexts, raising concerns about whether the reported improvements genuinely arise from subject-specific context adaptation.
  2. The methodological novelty of the proposed framework remains insufficiently substantiated. Several key components, including prototype-based retrieval and reliability-guided prediction correction, are closely related to existing TTA techniques. Although the authors discuss related approaches, the distinctive contributions of their integration are not adequately demonstrated. Furthermore, the lack of direct experimental comparisons with closely related reference-based TTA methods, such as T3A, TAST, and AdaNPC, makes it difficult to establish the proposed framework's methodological advantages over existing approaches.
Questions for authors
  1. Given the comparable performance obtained with source-derived and target-specific contexts in Appendix I, what evidence supports the necessity of subject-specific context for the proposed calibration mechanism?
  2. Can the authors clarify how frequently incomplete class coverage occurs in the evaluated target contexts and whether the proposed method yields greater improvements under such conditions?
  3. Can the authors provide paired statistical significance tests to assess the reliability of the improvements over NoAdapt, particularly on SEED-IV and FACED-9?
  4. How does the proposed framework differ from existing reference-based TTA approaches, such as T3A, TAST, and AdaNPC, in terms of its underlying adaptation mechanism? Can the authors provide comparable experimental results or clarify the limitations of such comparisons?
Reviewer_RPxcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NeuroContext-TTA for cross-subject EEG emotion recognition. It uses recent unlabeled target data to shift source prototypes, combines embedding and covariance retrieval, and applies gated corrections without updating model parameters at test time. The method shows improvements across four tasks, although the actual contribution of target context remains unclear. The method is evaluated on SEED, SEED-IV, FACED-3, and FACED-9 using five EEG backbones. On frozen mdJPT representations, it improves macro-F1 by 1.35, 0.54, 5.09, and 0.52 percentage points, respectively. The paper also provides component ablations, context-source interventions, feature perturbation analyses, and sample-level prediction diagnostics to investigate the behavior of the proposed calibration framework.

Critical strengths

The use of source prototypes to preserve all classes without relying on potentially unreliable target pseudo-labels. The evaluation covers multiple datasets and backbones, with useful ablations that help assess the method.

Critical weaknesses

My main concern is whether target context actually matters. On FACED-3, zero context achieves 36.10 F1, compared with 36.08 using real context. The results are also identical on SEED-IV and FACED-9. This makes the central motivation less convincing.

Questions for authors

Why does zero context perform almost identically to real target context? Could the authors provide a matched context-free baseline? Were target-subject statistics excluded from preprocessing? Could the authors also report paired confidence intervals for the improvements?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.