acceptodds
Under review as a conference paper at ICLR 2027

Subject-Coordinate Canonicalization: Learning Reusable Offsets for EEG Foundation Model Adaptation

Abstract

Electroencephalography (EEG) foundation models often generalize poorly to unseen subjects, but the structure of subject-dependent variation in their representations remains unclear. Across five EEG foundation models, we compare shared class structure with subject-specific translations and low-rank deformations. Held-out prototype and sample-level reconstruction analyses identify a substantial translation-like component, motivating adaptation through a reusable subject-level offset. Estimating this offset from unlabeled data is challenging because context statistics also reflect class composition. We propose Subject-Coordinate Canonicalization (SCC), which learns to estimate an offset from a small unlabeled context set and subtracts it from disjoint query representations. A permutation-invariant estimator aggregates sample-wise offset proposals, while class-balanced supervision on source subjects reduces sensitivity to context class composition. Leave-one-subject-out response alignment further encourages the corrected representations to preserve task-relevant structure shared across subjects. At inference, the estimated offset is cached and reused, requiring neither target labels nor gradient-based parameter updates. Experiments on three EEG benchmarks spanning emotion recognition, motor imagery, and sleep staging show improvements over source-only models and the evaluated adaptation baselines across all five foundation models. These results demonstrate that learned subject-level translation correction enables effective cross-subject adaptation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Foundation Models, Representation Learning, Cross-Subject Adaptation, EEG Decoding

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6Y5gOpenReview ↗
Rating
2
Confidence
3(−1)
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a test-time adaption method for EEG foundation models called Subject-Coordinate Canonicalization(SCC) which trains a subject-coordinate or subject-bias estimator on training set in units of episode (an episode contains multiple subjects and task classes so that model can learn to distinguish subject-dependent and task-related information).

This paper models latent representations produced by EEG-FMs as a composition of a subject coordinate, task-related information and residuals. Based on that, the estimator is trained to predict subject coordinate, which is defined as class-balanced subject center, obtained by first averaging the latent representations within each class and then averaging the resulting class means.

At the same time, the remaining part of the latent representations needs to align with shared task-related information, which is defined as the average task-related response of all other subjects for the same class. These two constraints encourage the estimator to learn the subject-coordinate or bias. They are denoted by LcoordL_{\text{coord}} and LcanL_{\text{can}}, respectively, in Equation 6. Beyond these losses, a shared classifier is trained on both the original and canonical representations, resulting in two additional losses, LbaseL_{\text{base}} and LtaskL_{\text{task}}, respectively.

The main contribution is the systematic diagnosis of latent representations produced by EEG-FMs, and it shows that a substantial component of cross-subject variation can be reduced by subtracting an estimated subject-related offset. Based on that finding, the author designed a pipeline to train an estimator that uses only a few context samples from the unseen subjects to predict the subject bias.

Critical strengths
  1. This work provides a systematic geometric diagnosis of EEG foundation model representations and models cross-subject variation in a simple manner, as subject-related information, task-related information, and residuals.

  2. Based on that, the authors propose a pipeline to train the subject coordinate estimator to reduce subject bias on unseen subjects, and this test-time adaptation method turned out to be effective compared with six baselines across three paradigms and three datasets.

Critical weaknesses
  1. Detailed experimental implementation information is missing. The number of repeated runs is not reported, and how the baseline TTA methods are applied is not reported in detail. The trainable parameter groups for Source-only, all baselines, and SCC are not clear. And since no code is provided, the results and training details remain unclear.

  2. Experiments are conducted only on three datasets, and those three datasets all contain a lot of subjects. However, many widely used EEG benchmarks, such as BCI Competition IV 2a, contain far fewer subjects. Hence, the robustness of this method when only a small number of source subjects is available remains unclear.

Questions for authors
  1. Can the authors evaluate SCC on a small-subject benchmark such as BCI Competition IV 2a, or otherwise provide evidence that it remains effective when only a small number of source subjects is available? This would clarify whether SCC remains effective when only a small number of source subjects is available or whether its performance depends on large and diverse source populations.

  2. Can the authors provide the complete implementation and evaluation protocol for Source-only, SCC, and every TTA baseline? In particular, please specify the trainable and frozen parameter groups, adaptation steps, optimizers, learning rates, hyperparameter selection procedures, number of repeated runs, and random seeds. Providing anonymized code and configuration files, if permitted, would further help verify the reproducibility and fairness of the comparisons.

Reviewer_XfwjOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a framework called Subject-Coordinate Canonicalization (SCC) which learns a subject-specific offset that is later used during inference to correct task-relevant structure and improve the model's performance.

Critical strengths

The method is based on a very-well known problem in EEG decoding. It attempts to solve inter-subject variability by looking it through the lens of component analysis (in a sense) and attributing the subject information to an offset that the model later uses to improve the performance of various tasks. I particularly liked the graphical contents and I do appreciate the effort that the authors put in various ablation studies (for example Table 2) which can definitely help the reader of this work.

Critical weaknesses

However, I think the method (although an interesting concept) lacks some fundamental work that needs to be completed by the authors.

  1. I find equation (3) a too simplistic decomposition of EEG signal analysis. How about the recorded device or site ? I am confident that these factors need to be part of the equation.

  2. The benefits of this methodology are very small. In most cases in Table 1 simpler methods like Euclidean Alignment or ReCenter get very close performance to the more complex and multi-step SCC proposed method.

  3. (Writing) I would recommend a re-structure of the paper. Currently, many important details are in the Appendix. In addition, there is a bad ordering in the content that makes the reader go back and forth. E.g. Appendix A is referred on first few pages and when the reader goes there, the variables that are used as not even defined until in later sections in the main text (and some are never defined e.g. equation (9)).

  4. (Missing Element) AI Statement is missing from the main manuscript (which was required in the Author Guidelines)

Questions for authors
  1. Is the SCC optimised for each dataset separately ? How about a cross-dataset analysis ? Or in datasets with multiple sessions of the same subject (get the offsets from one session and show the generalization in another) ? That would be a critical experiment to show if the method actually captures subject-related offsets effectively.

  2. In Table 1, are the results for the foundation models after full-fine tuning or probing ? Can you please also report the other results as well ?

  3. In Figure 5, why there is a drop ? Do the selected subjects matter ? How about an analysis of multiple seeds per step to get a mean and std.

Minor: Figure 1: Why is cache connected to the input EEG and not the b_{I} ?

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

This paper proposes Subject-Coordinate Canonicalization (SCC) for adapting EEG foundation models to unseen subjects using a small unlabeled context set. Motivated by prototype-level and sample-level analyses of subject-dependent representation shifts, SCC learns a permutation-invariant estimator that produces a reusable subject offset. This offset is subtracted from disjoint query representations. Training combines class-balanced coordinate supervision with leave-one-subject-out response alignment and additional regularization. At inference, the estimated offset is cached, requiring neither target labels nor gradient-based updates. Experiments cover five EEG foundation models and three datasets spanning emotion recognition, motor imagery, and sleep staging, with consistent improvements over the evaluated source-only and adaptation baselines.

Critical strengths
  1. The method is motivated by an explicit investigation of representation geometry. Comparing class-only, translation, and low-rank deformation hypotheses provides a useful empirical basis for the proposed adaptation mechanism. The additional sample-level analysis in Appendix A.5 strengthens this motivation by examining whether the observed structure persists beyond averaged subject-class prototypes.

  2. The context-query protocol is clearly defined and relevant to calibration with limited unlabeled data. Query samples are excluded from coordinate estimation, and the same estimated offset can be reused across subsequent predictions. Class-balanced source supervision addresses the important distinction between subject-dependent displacement and differences in context class composition.

  3. The experimental coverage is a substantial strength. Table 1 evaluates five backbones across three distinct decoding tasks, with SCC achieving the highest reported mean balanced accuracy and kappa in every backbone-dataset combination. The average balanced-accuracy improvements over source-only models are approximately 4.98, 4.03, and 3.74 percentage points on FACED, PhysioNet-MI, and ISRUC, respectively. This consistency makes the empirical contribution more convincing than an improvement limited to a single architecture or task.

  4. The practical inference procedure is simple: estimate one offset from a small context set, cache it, and apply subtraction to query representations. The component ablations and loss-weight sensitivity analyses provide additional evidence for the coordinate supervision and response-preservation objectives.

Overall, I lean toward acceptance. The combination of an interpretable adaptation mechanism, broad empirical evaluation, and low inference overhead constitutes a useful contribution. The concerns below mainly concern the scope of the interpretation and details needed to assess reproducibility.

Critical weaknesses
  1. The geometric conclusion should be qualified by the transformation family actually evaluated. The low-rank alternative in Eq. (10) uses subject-specific positive scaling within a shared subspace. It does not cover general subject-specific rotations, shears, or nonlinear transformations. Its limited benefit supports translation as an effective model within the tested family, but does not establish that other systematic deformations are generally negligible. The paper would be stronger if it consistently distinguished this empirical finding from a broader characterization of cross-subject EEG geometry.

  2. The explanation of context-size sensitivity is not fully established. Table 2 shows noticeable degradation when the random context size increases beyond four samples; for example, FACED balanced accuracy decreases from 62.63 at M = 4 to 60.25 at M = 32. Appendix E.3 provides a useful nested-context experiment, but taking nested subsets of the same pool does not by itself keep class proportions fixed. The procedure used to control composition therefore needs clarification. Reporting variation across repeated context draws with a fixed query set would help determine how much of the observed difference comes from context composition versus sampling variability.

  3. Some experimental details remain insufficiently specified. Table 1 reports standard deviations, and Figure 3 reports significance markers, but the number of runs, the source of the reported variability, and the statistical testing procedure are not clearly stated. The paper should also specify which encoder, readout, and classifier parameters are updated during source training and SCC training, and precisely identify the checkpoints used for the geometric diagnosis. These details matter for distinguishing the geometry present before SCC training from geometry encouraged by its objectives, and for understanding the contribution of the learned offset relative to source-stage representation learning.

Questions for authors
  1. Which checkpoints and representation layers are used for the analyses in Appendix A: pretrained models, source-only downstream models, or models trained with SCC objectives? Which components are frozen or updated at each stage?

  2. In Appendix E.3, how is context composition controlled when constructing nested subsets of different sizes? Are class proportions explicitly constrained, and are all samples in the maximum context pool excluded from the common query set? Can the authors report variability across repeated context draws?

  3. How many independent runs underlie Table 1, and do its standard deviations reflect training seeds, context sampling, or variation across subjects? What statistical test and unit of analysis are used for Figure 3? For the “w/o Coord. Estimator” ablation, which training losses and trainable modules are retained?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.