acceptodds
Under review as a conference paper at ICLR 2027

BiCSP: Learning Unified EEG Representations via Bidirectional Event–State Consistency

Abstract

Electroencephalography (EEG) exhibits neural dynamics at multiple levels of organization. Transient regional activity encodes fine-grained temporal events, whereas coordinated interactions across brain regions give rise to higher-level contextual brain states. Existing self-supervised EEG representation learning methods typically optimize either local temporal modeling or global contextual modeling, leaving the semantic relationship between these complementary scales largely unconstrained. We propose BiCSP, a hierarchical self-supervised framework for learning unified EEG representations through Bidirectional Cross-Scale Prediction. A learnable frequency-aware front-end first enhances neural signals in the spectral domain. Regional temporal patches are encoded into Event tokens, which are subsequently aggregated by a global attention encoder into a contextual State representation. BiCSP explicitly aligns these two representation levels through complementary latent prediction objectives: Event-to-State prediction reconstructs momentum-teacher State representations from Events, while State-to-Event prediction recovers masked Event targets from a context-isolated State together with visible Event tokens. Bidirectional local Event prediction and global-context masked prediction further regularize within-scale representation learning. After pretraining, Event and State embeddings are fused into a unified downstream representation and transferred using both linear probing and full fine-tuning. By explicitly enforcing semantic consistency across representation scales, BiCSP learns transferable EEG representations beyond conventional single-scale objectives.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, self-supervised learning, representation learning, neural signal processing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.60 over 5 reviews, above 10% of the others.
Reviewer_nDSZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper tries to model the relationship between local temporal events and global brain states. The main hypothesis is that bidirectional prediction between these two representation levels can improve transferable EEG representations. To validate this, they propose BiCSP, which combines frequency-aware processing, regional event encoding, global state aggregation, and 4 self-supervised prediction losses with an EMA teacher. The method then is evaluated on a limited number of EEG datasets using 2 protocals: linear probing and fine-tuning

Critical strengths

None

Critical weaknesses
  1. I have serious concerns about the experimental evaluation. The authors only compare with very old methods and miss many strong EEG FM backbones such as LaBraM [1], EEGPT [2], CBraMod [3], REVE [4], and CSBrain [5]. The authors discussed some of these works but never ran any experiments with them. Why don't the authors compare with these stronger baselines? Without these comparisons, I don't find any reason to support acceptance.

  2. The downstream evaluation is surprisingly limited. The authors only evaluate on five datasets, but most are related to cognitive states or emotion recognition tasks. There is no motor imagery BCI task such as BCIC-IV-2a, or sleep staging such as SleepEDF, which are common tasks for evaluating EEG FMs.

  3. The authors didn't sufficiently validate the hypothesis. The Global State is already obtained by aggregating Event tokens (Eq. 7), while Event-to-State prediction uses the same Event representations to predict a teacher State. For State-to-Event prediction, the predictor also receives visible Event tokens, not just the State. How can we know whether the improvement actually comes from learning bidirectional Event-State consistency? I find this nonsensical.

  4. There is no clear description of the pretraining setup: datasets used, # subjects and hours.

Questions for authors

See Weaknesses.

Reviewer_dNt9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

BiCSP learns hierarchical EEG representations using regional Event tokens and a segment-level State embedding. It combines a frequency-aware front-end, local and global predictive objectives, and bidirectional Event–State prediction with momentum-teacher targets. The representations are evaluated through linear probing and fine-tuning on five EEG benchmarks.

Critical strengths
  1. The paper does not clearly identify the pretraining corpus, its size, or whether a common pretrained checkpoint or separate dataset-specific checkpoints are evaluated. This prevents a precise interpretation of the claimed unified representation and transfer results.

  2. The manuscript reports five-seed means and standard deviations in the main tables and explicitly describes validation-based checkpoint and threshold selection.

Critical weaknesses
  1. The paper does not clearly identify the pretraining corpus, its size, or whether a common pretrained checkpoint or separate dataset-specific checkpoints are evaluated. This prevents a precise interpretation of the claimed unified representation and transfer results.

  2. Appendices B.1–B.2 provide general splitting principles, but omit dataset-specific partition sizes, subject assignments, and binary target definitions. Names such as “Crowdsourced” and “N-Back” are insufficient to identify the exact evaluation tasks.

  3. Section 3.1 says externally sourced results and official checkpoints are identified separately, but Tables 1–2 do not identify which entries fall into these categories. Consequently, readers cannot establish which comparisons share pretraining data, computational budgets, and reproduced implementations.

  4. Both global masked prediction and State-to-Event prediction use a State together with visible Events. The ablation labels do not clearly map to all four mathematical objectives, and no control establishes whether the predictor benefits from the State rather than relying primarily on visible Events.

  5. On STEW linear probing, BiCSP reaches 61.94% accuracy versus EEG2Rep’s 69.04%; on TUEV fine-tuning, its weighted F1 is 72.46% versus 75.08%. DREAMER accuracy gains are only 0.25 and 0.37 percentage points, without paired uncertainty analysis establishing their reliability.

  6. Figure 2 labels the regional Event encoder as “RNN feature extraction,” whereas Section 2.2 specifies a temporal Transformer. This concerns the encoder architecture itself and should be resolved rather than treated as a cosmetic labeling issue.

  7. The scope of the work and the rigor of its validation fall well short of the standard expected for publication at a highly regarded conference such as ICLR.

Questions for authors
  1. What exact data are used for pretraining, and are the five benchmarks evaluated using one shared checkpoint or separately pretrained models? Please specify the dataset identities, target labels, and partitions.

  2. Which Context Isolation variant produces the main results, what gap is used, and how do the two variants compare under identical conditions?

  3. What explains the differences between Figure 4’s clean accuracies and Table 1? Are the checkpoints, classifiers, and test partitions identical?

  4. Which losses are removed by each ablation label, and does State-to-Event prediction improve over an otherwise identical predictor receiving only visible Events?

Reviewer_1VcWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BiCSP, a self-supervised framework for EEG representation learning. A frequency-aware front-end and regional temporal encoder produce Event tokens, which an attention module aggregates into a global State. Training combines local temporal prediction, global masked prediction, and prediction in both directions between Events and States, using momentum-teacher targets. The downstream representation concatenates pooled Event features with the State. Evaluation covers five EEG benchmarks with linear probing and full fine-tuning, component ablations on four datasets, and a perturbation experiment on Crowdsourced.

Critical strengths
  1. The evaluation includes both frozen linear probing and full fine-tuning, and the main tables report means and standard deviations over five seeds. The strongest reported gains are on N-Back, where accuracy reaches 72.99% with linear probing and 77.31% with fine-tuning, compared with 62.30% and 63.33% for the best listed baselines.
  2. The component ablations provide some support for the design across multiple datasets. Under the same splits, optimization settings, and fine-tuning protocol, removing State-to-Event prediction reduces STEW accuracy from about 75.3% to 70.2% and DREAMER accuracy from 60.7% to 58.1%. Removing the frequency-aware front-end also lowers accuracy on all four datasets.
Critical weaknesses
  1. The main comparisons are not sufficiently documented. On DREAMER, Crowdsourced, STEW, and TUEV, every baseline mean in Tables 1 and 2 matches the corresponding value in Tables 2 and 3 of EEG2Rep, and all but two standard deviations also match. The submission does not identify which entries were reused and which were rerun, despite Section 3.1 stating that these sources are distinguished. The appendix gives general splitting rules but does not specify the exact train/validation/test partitions and task-label definitions for each dataset, or the pretraining data used in each experiment. These omissions make it difficult to establish whether the reported improvements were obtained under comparable training and evaluation conditions.
  2. The evidence does not clearly isolate the benefit of bidirectional Event–State prediction. Both the global objective and State-to-Event prediction recover masked teacher Events from a State and visible Events. Figure 3 does not clearly identify an Event-to-State ablation or test whether the State adds useful information beyond the visible Events supplied to the predictor. Appendix C.1 also describes masking before or after temporal self-attention without identifying which variant produced the main results. If masking occurs only afterward, retained tokens may already contain information from the excluded neighborhood. These uncertainties limit the evidence that the gains come from the proposed cross-scale relationship rather than additional masked prediction or access to local target information.
Questions for authors
  1. Which results in Tables 1 and 2 were evaluated for this submission, and which were quoted from prior work? For the newly evaluated results, which baselines used released checkpoints? Please specify the pretraining data, task-label definitions, and dataset partitions, and clarify whether preprocessing and model selection were matched. Was BiCSP pretrained separately for each benchmark or on pooled data?
  2. Which Context Isolation variant was used for the main results, and what masking ratio and isolation gap were used? Are results comparing the two variants in Appendix C.1 available?
  3. How does the global masked-prediction objective differ from State-to-Event prediction beyond Context Isolation, and which loss is removed in the “w/o State SSL” variant? Are results available for removing Event-to-State prediction alone, or removing the State input from the State-to-Event predictor while keeping the visible Event context fixed?
Reviewer_V3KPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BiCSP, a self-supervised learning framework for EEG data that explicitly links local temporal events with global contextual states through bidirectional cross-scale prediction. The method integrates a frequency-aware front-end, a hierarchical encoder, and a momentum teacher to jointly optimize the prediction of latent representations—mapping events to states and states to events—while fusing multi-level features for downstream tasks. Its core contribution lies in incorporating cross-scale semantic consistency into EEG representation learning. The authors evaluate the framework’s performance across five EEG datasets using linear probing, full fine-tuning, baseline comparisons, and experiments with five different random seeds; they further validate the contributions of individual components and the model's robustness through ablation studies and input perturbation experiments. The results demonstrate that this approach outperforms existing methods across most tasks and metrics.

Critical strengths

1.The research problem is clearly defined, and the methodology is underpinned by a coherent design rationale. To address the lack of explicit correlation between local temporal information and global context in self-supervised learning for EEG data, the paper proposes Event-to-State and State-to-Event which incorporates cross-scale relationships directly into the learning objective. This offers a clear and testable approach to the unified modeling of multi-level representations, which is the primary reason I find the methodological contribution of this work compelling.

2.The experiments are comprehensive, evaluating representation quality from multiple perspectives.

Critical weaknesses
  1. Context isolation may still allow prediction shortcuts. The original variant described in the appendix masks the target neighborhood only after temporal encoding, when target information may already have propagated to visible tokens through self-attention. No comparison with the pre-encoding isolation variant is reported, leaving it unclear whether this potential leakage contributes to the performance gains and weakens the proposed explanation for them.

  2. The State-to-Event predictor receives both the global State and visible Events, so it may complete the task using only the visible Events. Better prediction therefore does not establish that the model actually uses the State, weakening support for the central claim of bidirectional Event–State semantic consistency.

Questions for authors
  1. The paper emphasizes that bidirectional cross-scale prediction establishes Event–State semantic consistency, but the correspondence between the ablation settings and individual loss terms is unclear. Which loss does “w/o State SSL” in Figure 3 remove? Which experiments represent unidirectional prediction, bidirectional prediction, and training without cross-scale objectives? Are comparisons among Event-only, State-only, and fused representations available to assess their complementarity?

  2. Context isolation is well motivated, but the appendix includes a variant that applies isolation after encoding. Which variant was used in the main experiments, and what isolation gap was used? How do the two variants compare in performance? Please provide a direct comparison.

  3. The State-to-Event predictor receives both the State and visible Events. How can we establish that it actually uses the State rather than relying mainly on visible Events? Are there controls that remove or replace the State input to verify that reverse prediction constrains the learning of the global State?

Reviewer_5upfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BiCSP, a self-supervised framework for EEG representation learning that connects regional temporal Event tokens with a segment-level State representation. A frequency-aware front-end and temporal encoder produce Event tokens, which are aggregated through attention to form the State. Pretraining combines local bidirectional Event prediction, global-context masked prediction, and bidirectional Event–State prediction using momentum-teacher targets. State-to-Event prediction uses both a context-derived State and visible Event tokens, with a temporal neighborhood around each target excluded from the context. Pooled Event features and the State are concatenated for downstream classification. The evaluation covers five EEG benchmarks under frozen linear probing and full fine-tuning, alongside component ablations and an input-perturbation robustness experiment.

Critical strengths

The Event–State formulation offers a coherent way to connect regional temporal EEG features with a segment-level representation. The bidirectional latent-prediction objectives make this relationship an explicit learning target, providing a reasonable approach to studying how local and global representations can support one another. The evaluation covers five EEG benchmarks under both frozen linear probing and full fine-tuning. The ablations provide positive evidence for component utility: removing State-to-Event prediction lowers the reported accuracy on all four ablation datasets, by approximately 1.8–5.1 percentage points. These results support the usefulness of this objective within the complete framework.

Critical weaknesses

The added value of the Event–State formulation is not sufficiently established. The design is coherent and the component gains are encouraging, but the paper provides limited insight into what the proposed coupling adds to existing representation-learning approaches. This makes the methodological significance of the contribution less convincing. The experimental reporting leaves important uncertainty about reproducibility and interpretation. Key task and pretraining settings are insufficiently specified, and the relationship between the main evaluation and the robustness experiment remains unclear. These reporting gaps weaken the evidence supporting the paper’s conclusions.

Questions for authors
  1. Appendix C.1 describes two Context Isolation variants without specifying which is used in the main experiments. The masking ratio, isolation gap, and comparison results are also not clearly reported.
  2. Table 1 and Figure 4 report different clean accuracies for TF-C (64.82% vs. 74.9%) and MAEEG (75.21% vs. 65.0%). Both report Crowdsourced linear probing, but the relationship between these evaluations remains unexplained.
  3. The N-Back data source and labels, DREAMER target and labeling rule, and pretraining data are insufficiently specified. The relationship between pretraining data and downstream splits is also unclear, limiting reproducibility.
  4. Under the described pipeline, the spectral front-end (Eqs. (3)–(4)) can mix information from excluded neighborhoods into visible tokens before masking. The manuscript does not explain how Context Isolation handles this information path.
  5. Figure 2 labels the regional Event encoder as “RNN feature extraction,” whereas Section 2.2 and Appendix B.3 describe a two-layer Transformer. This is an unresolved inconsistency in the architecture description.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.