acceptodds
Under review as a conference paper at ICLR 2027

MART: Structured Regional Adaptation for EEG Representation Models

Abstract

Adapting EEG representation models to downstream tasks is often treated as a parameter-efficient update problem over the representation interface exposed by the backbone. We argue that, for EEG, the interface itself can be a limiting factor because flat hidden states may obscure the channel-time provenance and montage-dependent measurement conditions from which EEG representations are derived. We propose MART, a Measurement-Aware Regional Transformation framework that transforms this exposed interface into a structured regional adaptation space before lightweight downstream adaptation. Given EEG representations with channel-time provenance, MART recovers their channel-time organization, reparameterizes them into scalp-region representations, and augments each region with coverage and availability descriptors. A lightweight residual regional adapter and a linear prediction head are then optimized in this structured space. Across four downstream EEG tasks, MART achieves the best average Macro-F1 in our controlled adaptation comparison while updating only 0.108%–0.501% of the model parameters. Overall, MART reframes EEG representation adaptation as a problem of not only how to update a representation model, but also how its representations should be exposed to downstream adaptation.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
EEG representation learning, parameter-efficient adaptation, structured representation adaptation, montage-aware learning, regional token adaptation

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_baVROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The proposed method, MART, restructures frozen EEG representations with verified channel-time provenance into region-averaged tokens, appends regional coverage indicators, and trains a small adapter/head. It is tested and compared with LaBraM and REVE on DEAP, SEED, BCI IV 2a, and Sleep-EDF.

Critical strengths
  • Treating the adaptation interface, not only trainable parameter count, as a design variable.
  • The method sensibly avoids pretending pooled features have recoverable electrode identities, and includes missing-channel experiments.
Critical weaknesses

(1) There is a limited methodological originality and unclear contribution of the complete architecture. The central idea of reorganizing pretrained EEG representations into anatomically defined scalp regions before downstream adaptation is reasonable. However, the manuscript does not convincingly establish what new methodological principle is introduced beyond structured pooling and lightweight adaptation. The evaluation results (Table 3) show regional tokens with a linear classifier outperform full MART on DEAP and BCI IV 2a, using far fewer parameters. Thus, the evidence favors a simple regional readout, not the proposed full architecture. (2) There is insufficient evidence that anatomical regional structure is responsible for the improvements. The method assumes that aggregating pretrained representations according to anatomical scalp regions provides a better adaptation interface. However, regional averaging may suppress channel-specific information important for EEG decoding, for instance, spatially localized sensorimotor rhythms for motor imagery. The current experiments do not establish whether anatomical grouping is superior to other dimensionality-reduction or regularization strategies. A comparison with random channel grouping, spatially contiguous grouping, and data-driven grouping, under matched parameter budgets, would help determine whether the improvements arise from meaningful anatomical structure or simply from pooling and reduced dimensionality.

Questions for authors
  1. In Table 3, why does Regional Token + Linear outperform Full MART on both evaluated datasets? What is the specific benefit of the residual adapter?
  2. Does anatomical regional grouping outperform random grouping with matched region sizes and identical downstream architecture?
  3. How much of the robustness improvement is attributable to the coverage and availability indicators, rather than regional averaging alone?
  4. What are the exact subject/session splitting protocols for DEAP, SEED, BCI Competition IV 2a, and Sleep-EDF?
  5. Why do some task-specific EEG decoders outperform MART substantially, particularly on BCI Competition IV 2a, and what practical advantage does MART offer in these settings?
Reviewer_pXhsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MART, a lightweight regional adaptation for EEG representations.  Rather than relying solely on linear probing, MART constructs regional tokens by recovering channel-time structure from representations, aggregating temporal and spatial tokens within each brain region. These regional representations are then used for prediction, achieving comparable/superior performance to other parameter-efficient adaptation approaches.

Critical strengths
  1. The paper is generally well-structured and easy to follow, despite some unnecessary parts.
  2. The proposed method is straightforward and computationally efficient.
Critical weaknesses
  1. Limited technical contribution: MART appears to be an intuition-driven approach rather than a comprehensive technique. Its core is simple: reshape the representation to the signal’s geometry, then sum across temporal and spatial tokens to form a regional token. As such, I feel this work falls short of the technical standard of a top-tier conference like ICLR.
  2. MART lacks motivation/explanation for regional adaptation. The authors support regional embeddings without explaining: why we need channel-time-order embeddings rather than conventional EFM embeddings; or why we don't inject the task’s regional information at the beginning of downstream, as in CSBrain [1]. The authors should explicitly answer these questions before further structural design.

Minor concern: Overly detailed methodology and insufficient clarifications/explanations. The authors should merge/shorten several sections (Sec. 3.2, 3.3, 3.4, 3.5), as these components are straightforward. Further definitions and explanations of the baselines (Tab. 1, Tab. 2, Tab. 3) would be useful. Further explanation of the link between performance and the downstream task (e.g., why Linear-probe seems to work on Sleep but fails or performs worse on Motor Imagery (BCI IV 2a)) would provide useful insight.

[1] CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding, 2026

Questions for authors

Weaknesses above

A common motivation for EEG foundation models is to learn generalizable representations that encode spatial, temporal, and semantic information for adaptation. As such, why does recovering their channel-time arrangement provide additional information? Please prove this to strengthen the work.

Reviewer_GD6fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes MART, a Measurement-Aware Regional Transformation framework that transforms this exposed interface into a structured regional adapta-tion space before lightweight downstream adaptation. Given EEG representationswith channel-time provenance, MART recovers their channel-time organization,reparameterizes them into scalp-region representations, and augments each re-gion with coverage and availability descriptors. A lightweight residual regionaladapter and a linear prediction head are then optimized in this structured space. Across four downstream EEG tasks, MART achieves the best average Macro-F1 in the controlled adaptation comparison while updating only 0.108%-0.501% of the model parameters. Overall, MART reframes EEG representation adaptationas a problem of not only how to update a representation model, but also how itsrepresentations should be exposed to downstream adaptation.

Critical strengths

1.The paper highlights an important aspect of EEG foundation model adaptation: downstream performance depends not only on how model parameters are updated, but also on how pretrained representations are organized. By recovering channel-time structure, constructing anatomical regional representations, and incorporating regional measurement information, MART provides a meaningful perspective on structure-aware and parameter-efficient EEG adaptation. 2.The paper evaluates MART across four EEG downstream tasks and compares it with a range of parameter-efficient adaptation methods. The results demonstrate competitive performance with relatively few trainable parameters. Furthermore, the capacity-controlled experiments, channel-dropping robustness evaluation, and regional sensitivity analysis provide useful insights into the effectiveness of structured regional representations.

Critical weaknesses

1.MART relies on predefined anatomical regions determined by dataset-specific electrode montages, while different EEG downstream tasks may depend on distinct spatial patterns. The paper does not systematically evaluate alternative regional partitioning strategies or demonstrate that the predefined regions adequately capture task-relevant information. Although Figure 3 reports regional sensitivity under structured region removal, these results do not establish the suitability of the chosen partitions. Consequently, it remains unclear whether the improvements arise from meaningful anatomical structure or generic feature grouping and aggregation. 2.The paper lacks a systematic analysis of when structured regional adaptation is beneficial. Table 1 shows that MART underperforms conventional Adapter on LaBraM with SEED and REVE with Sleep-EDF, while Table 3 indicates that the simpler Regional Token + Linear variant outperforms Full MART on DEAP and BCI IV 2a. These results suggest that the effectiveness of regional representations and the additional adaptation components may depend on specific tasks and data characteristics. However, the paper does not sufficiently investigate when regional aggregation improves downstream prediction, whether it may discard task-relevant channel-time information, or under what conditions Full MART provides advantages over simpler regional approaches. Consequently, the applicability of the proposed method and the necessity of its complete architecture remain insufficiently established.

Questions for authors

1.Table 1 shows that MART outperforms conventional adaptation methods on several tasks but underperforms Adapter on LaBraM with SEED and REVE with Sleep-EDF. Could the authors further explain these performance differences? In particular, what task or data characteristics make structured regional representations beneficial, and could predefined regional partitioning and channel-time aggregation discard fine-grained information important for certain tasks? 2.Table 3 shows that Regional Token + Linear outperforms Full MART on both DEAP and BCI IV 2a despite using substantially fewer trainable parameters. Could the authors explain why the additional residual regional adapter does not consistently improve performance? Furthermore, under what conditions does Full MART provide advantages over simpler regional linear approaches, and what are the respective contributions of coverage encoding, availability encoding, and residual adaptation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.