- 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
- Avoiding target-time optimization addresses a practical problem: fitting a large model to a few target trials can be unstable.
- 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
-
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.
-
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.
-
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.
-
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.
-
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.
-
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
-
What source training, initialization, auxiliary supervision, and validation-based tuning does each baseline receive?
-
Can you report comparisons from a matched source-trained checkpoint and evaluate the support encoder's standalone prototype classifier?
-
What explains the zero-shot and fine-tuning differences between Table 1, Figure 3, and Figure 4?
-
How do you interpret the 62.0% label-shuffled result relative to 62.5% with true labels?