acceptodds
Under review as a conference paper at ICLR 2027

SEAM: Support-Encoded Adaptation with Meta-Learned Prompting for Few-shot Cross-Subject EEG Foundation Model Adaptation

Abstract

Recent advances in deep learning have improved EEG decoding by learning task-relevant representations from neural recordings. Large-scale pretraining has further enabled EEG foundation models that capture reusable structure across subjects and tasks. However, adapting these models to an unseen subject from only a few labeled trials remains unreliable. Fine-tuning on small, subject-biased support sets can distort pretrained representations and induce negative transfer. We propose SEAM (Support-Encoded Adaptation with Meta-Learned Prompting), a target-time optimization-free framework for few-shot cross-subject EEG foundation model adaptation. SEAM jointly meta-trains the Transformer backbone, classifier head, and a lightweight prompt generator on source subjects, then freezes their weights for target adaptation. From a target support set, it constructs a global subject prototype and class residual prototypes; each query routes over these residuals to generate a personalized prompt. This preserves task-generic knowledge while expressing subject- and query-specific variations without target-time updates. Experiments on different EEG benchmarks show that SEAM better exploits limited support data, mitigates negative transfer from 2.4% to -4.8% and update-induced forgetting, and achieves a favorable accuracy–efficiency trade-off.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG Foundation Model, Few-shot Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_rFJoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies adaptation of an EEG foundation model to a new subject using a small labeled support set. SEAM jointly meta-trains a BrainOmni backbone, classifier, and support encoder on source subjects. At deployment, all weights are fixed. A global support prototype and query-dependent routing over class residual prototypes generate input prompts. Experiments on PhysioNet-MI and FACED compare this approach with fine-tuning, parameter-efficient adaptation, and few-shot baselines, with additional prompt controls and component ablations.

Critical strengths
  1. Avoiding target-time optimization addresses a practical problem: fitting a large model to a few target trials can be unstable.
  2. The prompt replacement controls in Section 5.6 provide useful evidence that target support information matters. Real support prompts achieve 62.5% accuracy, compared with 57.3% for train-mean prompts and 55.7% for subject-shuffled prompts.
Critical weaknesses
  1. The experiments do not isolate the contribution of SEAM's adaptation mechanism. SEAM updates the backbone and classifier during source meta-training and introduces a separate raw-EEG encoder with prototype supervision. The manuscript does not establish that the baselines receive comparable source training, initialization, or validation-based tuning. Improvements could therefore come from additional source training or the auxiliary encoder.

  2. Unexplained differences between reference results make the comparisons difficult to interpret.​ For PhysioNet-MI at 4-shot, Table 1 reports 59.8% zero-shot accuracy, whereas Section 5.6 and Figure 4 report 41.2%. Figure 3 gives target accuracies of 29.9% and 47.9% for head-only and full fine-tuning, compared with 52.1% and 58.0% in Table 1. Different checkpoints or evaluation protocols could explain these differences, but the manuscript does not identify them clearly.

  3. The controls provide limited evidence for the proposed class-residual routing mechanism. Figure 4 reports 62.0% accuracy after shuffling support labels, compared with 62.5% using the true labels. This control appears in the figure but is not discussed in the accompanying text. Table 3 shows similarly small accuracy reductions when query conditioning or the support-prototype branch is removed: 0.5 and 0.2 percentage points, respectively. These results leave open whether most of the benefit comes from label-free subject-level conditioning. A global-prototype-only prompt is a direct control for that explanation.

  4. Reliable mitigation of negative transfer has not been demonstrated across the reported settings. On FACED at 2-shot, SEAM achieves 51.4% accuracy, below the 51.8% zero-shot baseline. On PhysioNet-MI at 2-shot, the gain is only 0.1 percentage point. The tables provide no uncertainty across subjects, training seeds, or support draws, making the reliability of these differences unclear.

  5. The forgetting analysis does not establish preservation of general foundation-model capabilities.​ Freezing target-time weights guarantees that target adaptation cannot overwrite those weights. However, SEAM changes the backbone during source meta-training, and its zero-shot loss supervises the downstream source task without directly constraining retention of the original pretrained behavior. The reported increase from 47.4% to 70.2% on held-out source data compares unprompted with prompted inference, introducing support information into the comparison. It therefore cannot isolate knowledge preservation.

  6. The claimed accuracy-efficiency advantage lacks quantitative support.​ The manuscript reports no target-time latency, peak memory, or parameter overhead under matched hardware and batch sizes. Avoiding gradient updates removes an adaptation cost, but SEAM adds support encoding, query-dependent prompt generation, and additional transformer tokens. These costs should be included in comparisons with fine-tuning and parameter-efficient alternatives. Reporting source meta-training cost alongside deployment savings would also clarify when the additional training is amortized.

Questions for authors
  1. What source training, initialization, auxiliary supervision, and validation-based tuning does each baseline receive?

  2. Can you report comparisons from a matched source-trained checkpoint and evaluate the support encoder's standalone prototype classifier?

  3. What explains the zero-shot and fine-tuning differences between Table 1, Figure 3, and Figure 4?

  4. How do you interpret the 62.0% label-shuffled result relative to 62.5% with true labels?

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

This paper proposes SEAM: Support-Encoded Adaptation with Meta-Learned Prompting for Few– shot Cross-Subject EEG Foundation Model Adaptation. This paper discusses an interesting topic. But, from my perspective, generally this paper is not ready yet for a publication. Please see weaknesses.

Critical strengths
  1. This paper discusses an interesting and applied topic
Critical weaknesses
  1. Problem formulation: This paper has an unclear objective, what kind of problem does the author want to solve? Is it domain adaptation or few-shot learning or few-shot domain adaptation or few-shot continual learning, or any other problem? It should be clearly stated and formulated.
  2. Proposed Methodology: It claims to propose a meta-learning-based algorithm, but the objectives of both the base learner and the meta learner are unclear. The interaction, or how the meta-learner guides the base learner, is unclear.
  3. Novelty: Looking at section 4, I cannot find a sufficient novelty of the proposed method. On the contrary, the proposed methods are very much basic.
  4. Unclear evaluation setting: I do not believe that a common inference phase requires a support set and query set as shown in figure 2.b. If yes, how can you confirm its practicability?
  5. Unclear baselines: Tables 1 and 2 show confusing results and analysis. What problem setting is exactly what the proposed method/model wants to solve? Is it few-shot classification or domain adaptation, or other settings? In addition, the baselines included in the tables are all very common baselines for the respective domains that are usually included as an upper/lower bound.
  6. Section 5.5 discusses forgetting, but Figure 3 and the analysis clearly discuss accuracy. Again, please define your problem, objective, and evaluation settings properly before conducting the experiment and analysis.
  7. The writing lacks substantial details while leaving much unfilled space of the page limit. This indicates that the paper is not ready.
Questions for authors

Please address weaknesses.

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

The proposed SEAM framework jointly meta-trains an EEG foundation-model backbone, classifier, and support-conditioned prompt generator on source subjects. At deployment, all model parameters are frozen. Support examples produce a global prototype and class residual prototypes; these are combined with each query to generate prompt tokens.

Critical strengths

The task is practically relevant: fine-tuning on very small support data sets can be unstable. Eliminating additional optimization simplifies deployment, and the distinction between source training and target inference is stated clearly.

Evaluation performed on both motor-imagery and affective EEG tasks is useful, as is reporting macro F1 and kappa alongside accuracy. The subject shuffled and mean prompt controls also provide evidence that target specific support information matters.

Critical weaknesses

Result reported in different places for PhysioNet-MI 4-shot differ substantially, e.g. 59.8% for zero shot in table 1 while 41.2% in Figure 4. 52.1% for head fine-tuning in table 1 while 29.9% in Figure 3. The paper does not make it clear the reason behind such performance variations.

According to Section 4.2, a consistent permutation of support-class labels merely permutes the prototypes and their routing weights. Their weighted sum remains unchanged. The routing mechanism discards the identities attached to support-class prototypes in this case.

In Table 3, removing query conditioning changes accuracy from 62.5% to 62.0%, while removing the support prototype branch changes it to 62.3%. Removing global context causes a much larger drop, to 56.8%. It becomes unclear whether it is necessary to have the mechanisms of query conditioning and support prototype branch in this case, as they don't influence performance much.

Questions for authors

Please see above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.