acceptodds
Withdrawn from ICLR 2027

Can Cross-Identity Session-Shift Transfer Improve EEG Identification Across Sessions?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-session EEG identification requires identity-discriminative representations to remain stable despite changes in cognitive state and acquisition conditions. We propose SHIFT-ID, a cross-identity session-shift transfer framework for the setting in which registered users provide no calibration data in their later evaluation sessions.SHIFT-ID constructs an empirical drift bank from paired sessions of auxiliary identities in the space of feature means and log standard deviations. It then transfers drifts across identities to generate session-informed training views and combines cosine consistency between representations with worst-view prototype regularization to preserve identity structure under these perturbations. At inference time, the fixed encoder, projection head, and classifier process each query independently, without calibration recordings from evaluation users or online parameter updates. We establish a unified information-isolation protocol on MCD, M3CV, and SEED-IV, evaluate twelve commonly used or recent EEG architectures, and perform information-budget-matched comparisons, transfer-direction analyses, mechanism ablations, and perturbation diagnostics. All experiments are repeated five times, and performance is reported as the mean over the five runs. The full configuration achieves mean Rank-1 accuracies of 70.94%, 77.83%, and 85.06% on the three datasets, respectively. Relative to CAGCNet, the strongest source-only external baseline under our protocol, the corresponding gains are 0.55, 0.42, and 0.38 percentage points. Under the same backbone and auxiliary-target budget, the Rank-1 gains over DANN are 0.63, 0.31, and 0.27 percentage points. Across the evaluated conditions, the directional, ablation, and perturbation analyses consistently support the contribution of empirical drift transfer and robust prototype regularization, while the modest absolute margins motivate a conservative interpretation of the improvements.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG identity recognition, cross-session transfer, feature statistics, robust prototypes, calibration-free recognition

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_T8hpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors investigate EEG-based person identification across recording sessions. At training time, recordings from a first session are available for all enrolled users, but recordings from a later session are available only for a subset of them, referred to as auxiliary users; the remaining users are evaluated on their later session without any calibration data. The proposed method, SHIFT-ID, measures for each auxiliary user how the statistics of the learned features (channel means and log standard deviations) change between the two sessions. These session shifts are then applied to the features of other users to simulate how their recordings might change in a later session, and the network is trained to give consistent identity predictions on these simulated views, using a consistency loss and a worst-case prototype loss. The method is evaluated on three datasets (MCD, M3CV, and SEED-IV) against twelve EEG architectures and against strategies that use the same auxiliary data, such as DANN.

Critical strengths

First, the information-isolation protocol is clearly specified, and the comparison group that uses the same backbone and auxiliary data is the right design for separating the value of the data from the value of the method.

Second, the mechanism is simple, its components are examined through ablations, and the authors report the modest size of the gains candidly.

Critical weaknesses

First, the gain is small and not statistically supported. Around 11 points on MCD come from the additional target-session data, and SHIFT-ID outperforms both cross-entropy on that data and DANN by less than one point. With 5–8 evaluation identities on two datasets and no standard deviations or statistical tests, it is not possible to tell whether these improvements exceed run-to-run variability, although the paired design and the consistent direction of the gains help. Moreover, CAGCNet without the auxiliary data already matches the authors’ backbone trained with it, suggesting the architecture matters more than the training method.

Second, the novelty is limited, and the authors do not explain their hyperparameter choices. Transferring within-class variation to other classes has been shown before (Hariharan and Girshick, 2017; Schwartz et al., 2018), and feature-statistic augmentation is close to AdaIN, MixStyle, and DSU, so the contribution lies mainly in applying this idea to session differences for EEG identification. The default matches the best test result in every MCD sweep, but the paper does not explain how the defaults were chosen, which is concerning when the gains are under 1 point. On presentation, C0–C8 in Table 3 are not defined, and Appendix A largely repeats Section 3.

Questions for authors

Can the authors provide standard deviations and a paired test across multiple runs, or an identity-level bootstrap interval, for the comparisons with auxiliary cross-entropy, DANN, and CAGCNet? How were the default hyperparameters chosen? It would also be useful to see SHIFT-ID applied to the strongest backbones with the same auxiliary data, and compared with an augmentation that uses the auxiliary target statistics, such as MixStyle. What is the effect of combining SHIFT-ID with auxiliary cross-entropy? Regarding the data, were the two MCD conditions recorded on different days? How does the gain change as the number of auxiliary identities increases on M3CV?

Reviewer_Lm5LOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SHIFT-ID addresses closed-set EEG identification when source-session recordings exist for all enrolled identities, but paired target-session recordings are available only for a disjoint auxiliary subset. After source-only warm-up, it estimates each auxiliary identity's session change in feature means and log standard deviations. These fixed drift vectors perturb source features from other identities. Training combines original-view identity classification, cosine consistency, and a worst-view in-batch prototype loss. Inference uses only the fixed encoder, projection head, and classifier.

The paper evaluates MCD, M3CV, and SEED-IV, including source-only architectures, methods with the same backbone and auxiliary-target budget, transfer directions, perturbations, and ablations. Five-run mean Rank-1 accuracies are 70.94%, 77.83%, and 85.06%. Improvements over matched DANN are 0.63, 0.31, and 0.27 percentage points. The setting includes auxiliary target data and should be understood as closed-set recognition with auxiliary histories, not source-only generalization or open-set authentication.

Critical strengths
  1. The paper treats information access carefully. Evaluation-user target recordings are excluded from training, normalization, model selection, and test-time adaptation, while predictions retain the full enrolled candidate set. Table 1 separates source-only architectures from strategies with the same backbone and auxiliary-target budget. The text explicitly acknowledges that the large gain over source-only CE includes additional information. These distinctions make the evaluation much more interpretable.

  2. The design is simple at inference, and the diagnostics address several relevant mechanisms. Drift sampling and prototypes are training-only components. The experiments cover three datasets, both MCD directions, all six SEED-IV directions, auxiliary-data budgets, prototype self-inclusion, and drift-bank updates.

Critical weaknesses
  1. Training and evaluation remain small in scale. The datasets contain only 15, 23, and 95 subjects, with fixed auxiliary/evaluation splits. EEG corpora with thousands of subjects exist, such as TUH EEG. The experiments do not establish whether the method remains useful with larger-scale pretraining or identification among substantially more people.

  2. The novelty relative to CAGCNet needs clearer justification. CAGCNet already combines feature augmentation and contrastive learning for cross-session identification. The paper does not clearly establish the added value of this design over CAGCNet given the same data access.

  3. Gains are very small and lack uncertainty estimates. Improvements over DANN with the same backbone and data are only 0.63/0.31/0.27 percentage points. CAGCNet is also close despite using source data only (Table 1). Five-run means without standard deviations or confidence intervals do not establish that these gains are reliable.

Questions for authors
  1. Do the gains persist with larger training populations, including broader EEG pretraining, and larger identification candidate sets?

  2. What is technically new relative to CAGCNet, and does SHIFT-ID outperform it when both have access to the same auxiliary data?

  3. Can you report uncertainty for the gains over matched baselines and test their robustness across different auxiliary/evaluation subject splits?

Reviewer_wYhrOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies cross-session EEG identification when target-session data from evaluation identities are unavailable for calibration, while paired historical sessions from auxiliary identities can be used. The authors propose SHIFT-ID, which estimates session drifts from feature means and log standard deviations of auxiliary identities and transfers these drifts to generate augmented views for other identities. The method combines cosine consistency and worst-view prototype regularization to preserve identity-discriminative representations. The authors evaluate the method on MCD, M3CV, and SEED-IV, comparing against twelve external EEG architectures and several information-matched training strategies. The full method achieves Rank-1 accuracies of 70.94%, 77.83%, and 85.06%, respectively. Although the problem setting is meaningful and the paper provides extensive experiments with positive mean results, the specific scientific value of transferring identity-level session drifts remains insufficiently established. Together with the small and statistically under-supported improvements, this substantially weakens the claimed contribution and the confidence in the empirical conclusions.

Critical strengths
  1. A meaningful evaluation setting with clear information constraints. The paper separates auxiliary identities from evaluation identities, while keeping all enrolled identities in the final classification space (Section 3.1 and Appendix B.2). Evaluation target-session data are excluded from training, normalization, and model selection. This is a reasonable setting for studying cross-session identification without user-specific recalibration. I also appreciate that the authors distinguish this setting from source-only domain generalization and unseen-identity recognition.

  2. The comparison design considers differences in training information. Table 1 includes twelve external EEG architectures and several same-backbone strategies, including Source+Aux CE, Deep CORAL, DANN, and RandomShift. Appendix B.3 also describes shared warm-up weights and training seeds for same-backbone comparisons. Importantly, the authors distinguish source-only methods from those using auxiliary target-session data, and acknowledge that the large improvement over Backbone CE cannot be attributed entirely to the proposed mechanism. This makes the experimental comparisons more careful.

  3. The paper includes several specific mechanism diagnostics and relatively careful discussions. Table 2 evaluates both MCD transfer directions, all six SEED-IV directions, and input perturbations. Table 3 further examines drift components, pairing and coordinate structure, prototype construction, drift-bank updates, and hyperparameter sensitivity. The auxiliary-data budget experiments are also useful. These analyses provide some supporting evidence, and the authors explicitly acknowledge the small performance margins, task-session confounding, and limitations of five-run means.

Critical weaknesses
  1. The novelty and specific value of identity-level drift transfer are not sufficiently established. The proposed feature-statistic transformation is closely related to existing augmentation methods, while consistency and prototype regularization build on established ideas (Equations 2–7). The key contribution therefore lies in how auxiliary identity drifts are estimated and transferred. Although the authors provide pairing and randomized-coordinate ablations (Table 3), there is no comparison with a simpler global mean session drift or other information-matched statistical augmentation. MixStyle is only evaluated in the source-only group (Table 1). Stronger evidence is needed to establish the independent scientific value of the proposed drift structure.

  2. The performance improvements are small, with insufficient statistical support. In Table 1, SHIFT+RP improves Rank-1 over the information-matched DANN by only 0.63/0.31/0.27 percentage points, and over SHIFT by 0.28/0.22/0.19 points. Although experiments are repeated five times, only means are reported, without paired uncertainty estimates or per-run results (Section 4, Appendix B.4). The identity partitions are also fixed, with only 8 and 5 evaluation identities on MCD and SEED-IV. These results make it difficult to judge whether the small gains are reliable across training runs and identity splits.

  3. The independent contribution of worst-view prototype regularization is still unclear. Equation 5 includes the query and its augmented views in prototype construction. Although P0–P2 examine this issue, they use a different batch-sampling scheme, without a corresponding no-RP baseline (Table 3). Also, worst-view aggregation improves over average-view aggregation by only 0.20/0.16/0.14 points (Section 6). Since the regularizers operate on normalized embeddings while the final classifier uses unnormalized embeddings (Equations 4–7), a direct comparison with shifted-view classification or prediction-consistency objectives would better establish the value of the proposed regularization.

  4. Some ablation and hyperparameter-selection details are unclear. Table 3, Panel A uses C0–C8 without explicitly defining every configuration, making several ablation results difficult to interpret. In Panel B, the default hyperparameters achieve the highest reported MCD test Rank-1 across the sensitivity groups. Although Appendix E.2 states that evaluation-target data were not used for hyperparameter selection, the actual selection data and criteria are not sufficiently described. These details matter given the small reported improvements.

  5. Some protocol and reproducibility details need clarification, and the figures require further polishing. Appendix B.2 describes a stratified source split on MCD and mentions recording-level grouping, but the actual grouping procedure is not fully clear. The paper provides a data availability statement, but I could not find the implementation or complete experimental scripts. Figure 1 is dense, and several Figure 2 panels have crowded labels. Figure 2(k) also needs clearer details about identity selection, dimensionality reduction, and the colored regions. These issues make the results harder to interpret and reproduce.

Questions for authors
  1. Could the authors compare the identity-level drift bank with a global mean session drift or a distribution-matched perturbation under the same auxiliary-data and training budget? This would help clarify the specific benefit of transferring individual auxiliary drifts.

  2. Could the authors provide paired per-run differences and uncertainty estimates, especially against DANN and SHIFT? Results across additional auxiliary/evaluation identity splits would also help establish the stability of these small improvements.

  3. Could the authors evaluate RP against a no-RP baseline under the same batch sampling used for P0–P2? How does the proposed regularization compare with directly enforcing classification or prediction consistency on shifted views?

  4. Could the authors explicitly define C0–C8 in Table 3 and clarify the actual recording-level splitting and hyperparameter-selection procedures? These details would make the experimental evidence easier to assess.

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

Related papers

Discussion (0)

Sign in to comment.