- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper studies whether attention weights in motor-imagery EEG decoders contribute to predictions in a trial-specific way. The authors replace the weights at frozen checkpoints with weights from other trials or with training-set averages, while keeping the current trial’s Value tensors and model parameters fixed. They find that several existing decoders are relatively insensitive to these replacements and discuss how token similarity, the Value representation, and downstream processing can account for this behavior. They then propose SOAMNet to make predictions depend more strongly on sample-conditioned attention. Across four datasets, they report competitive classification performance, improvements over a matched non-attention head, and substantially larger performance drops when SOAMNet’s attention weights are replaced. The paper presents these results as a diagnosis and repair of attention failure.
Critical strengths
-
The replacement audit makes a useful distinction between the performance of a complete decoder and its reliance on the current trial’s attention weights. Holding the Value tensors and parameters fixed is a reasonable way to examine that distinction, and the results caution against inferring functional importance from attention maps or overall accuracy alone.
-
The authors provide substantial empirical analysis, including classification comparisons, replacement interventions, and controls that examine the source of SOAMNet’s sensitivity. I particularly appreciate the global-context controls in Table S19: they provide evidence that is important for interpreting what the proposed model actually does. These are useful contributions, even though I do not think they support the central repair claim.
Critical weaknesses
-
The paper does not establish that replaceable attention weights constitute a functional failure. The audit shows that, at these checkpoints, replacing the current trial’s attention weights with cross-sample or training-set-mean weights changes balanced accuracy only slightly. This shows that the trial-specific weights are not necessary for these predictions, but it does not show that the decoders are deficient. If the class-relevant information is already captured by global statistics, or remains available through residual or downstream paths, low sensitivity to replacement may reflect a reasonable solution to the task, redundancy, or robustness. To call this a failure, the paper would need to show what is lost because of it, for example in accuracy, generalization, robustness, or a clearly specified faithfulness requirement. Without such a criterion, the target of the repair is defined by the expectation that attention should be used. SOAMNet is then designed to force evidence through one sample-conditioned attention path, and the large drops under replacement confirm that its predictions depend on that path. However, greater sensitivity to replacing an internal component is not, by itself, evidence of a better model, and could equally indicate fragility. This is the main reason for my recommendation, because the diagnosis-and-repair framing organizes the whole paper.
-
The paper’s own controls suggest that SOAMNet’s dependence comes mainly from a global class-conditioned pathway, rather than from better selection of temporal evidence. SOAMNet first forms a global class context and writes it into the attention scores through a fixed temporal carrier. In this design, attention can carry class information to the output without needing to identify which EEG time points are informative. Swapping attention across trials then partly amounts to attaching one trial’s class-related code to another trial’s evidence, so a large drop in performance is expected. Table S19 is consistent with this interpretation: the global-carrier-only model retains core replacement losses close to those of the full model (25.38/41.22/32.23/20.61 vs. 28.46/40.95/33.81/18.61 pp on 2b/2a/BNCI/Cho), whereas removing the content Q–K scores costs only 1.08–4.29 pp, and swapping attention after matching global evidence costs only 1.32–3.12 pp. The authors acknowledge the role of the global context, but the paper still presents the large dependence as evidence that attention failure has been repaired. A more accurate conclusion would be that the strong dependence arises mainly from a class-conditioned global pathway, while trial-specific temporal content contributes a small additional effect. The accuracy gains over the baselines and over the matched non-attention head are a practical contribution, but they do not show that these gains come from repairing attention failure, or that strong attention dependence is what produces them. Addressing this would require reframing the central claim, not only adding experiments.
Questions for authors
-
Beyond the replacement results themselves, what concrete cost do you attribute to replaceable attention in existing decoders? Is there evidence that decoders with low replacement sensitivity perform worse in accuracy, cross-session or cross-subject generalization, robustness, or another clearly specified criterion? Evidence of such a cost would most directly affect my assessment, because it would give the diagnosis-and-repair framing an independent basis.
-
Can the gains from the global class-conditioned pathway be separated from those of trial-specific temporal selection? For example, how does an independently trained model with matched capacity and information perform when the same global class context is provided directly to the classifier, or through a simple non-attention gate? If such a model matches SOAMNet’s accuracy without strong attention dependence, how would you interpret the repair claim?
-
Clarification: please describe the matched non-attention head in more detail, including its inputs, whether it has access to the global class context, its parameter count, and its training procedure. This would make clear which comparison the reported 3.51–5.22 point difference reflects.
4 comments
Thank you for clarifying why mechanistic misattribution matters and how the pooling baseline was constructed. My main concern remains whether the reported classification gains can be attributed to the mechanism the paper claims to repair. The comparison shows that SOAMNet outperforms a pooling head given the same source trajectories, but it leaves open whether a direct readout given the same computed global class context could achieve similar accuracy. This distinction matters because the global-carrier-only model retains most of the replacement sensitivity, and adding a bypass largely eliminates that sensitivity with a much smaller change in clean accuracy. The response establishes that SOAMNet uses trial-specific routing; the evidence still leaves open what that routing adds beyond the construction and use of global class evidence.
Routing Utility Beyond Global Class Evidence
Thank you for the follow-up.
We agree that this is a valid concern, so we added a control that passes the global class context computed by SOAMNet directly to an independently trained classification head. SOAMNet outperforms this control on all four datasets, showing that attention routing provides additional classification gains beyond direct use of the same global class information.
For each participant and random seed, we froze the trained SOAMNet encoder, evidence aggregator, and global-context computation. The control's classification head uses only the global class context computed by SOAMNet before routing. It receives no attention scores, attention weights, Values, local temporal trajectories, or routed outputs.
We independently trained a three-layer residual MLP as the control's classification head. The control model's total active parameter count differs from SOAMNet's by no more than 1%. Both models use the same training, validation, and test splits, three random seeds, and training budget. Checkpoints are selected by validation BA for both models.
After averaging the three seeds within each participant, the test results are:
| Dataset | SOAMNet BA | Direct-readout BA | Paired difference [95% CI] |
|---|---|---|---|
| BCIC-IV-2b | 80.10 | 77.75 | +2.35 [1.35, 3.35] pp |
| BCIC-IV-2a | 69.44 | 68.19 | +1.25 [0.60, 1.90] pp |
| BNCI2015-001 | 85.53 | 83.98 | +1.55 [0.85, 2.25] pp |
| Cho2017 | 72.75 | 69.30 | +3.45 [2.72, 4.18] pp |
SOAMNet achieves higher BA with each of the three random seeds and for 72 of the 78 participants. Cohen's also improves on all four datasets, with the confidence intervals for the paired differences excluding zero.
The earlier global-carrier and bypass controls showed that global context carries substantial class information. The new control uses that information directly, yet its BA remains 1.25--3.45 percentage points below SOAMNet's. These results show that attention routing can improve classification beyond direct use of the same global class information. We will include this control and its results in the revised manuscript.
Follow-up on Our Response
If you have had a chance to read our follow-up response, we would appreciate hearing your thoughts. Thank you for your time.
Evidence for Attention Failure and Functional Route Repair
Thank you for the detailed review and comments.
Response to Critical Weakness 1 and Question 1: criterion and cost of attention failure
The attention failure identified in the paper is clearly established. We must start from a basic fact: an effective attention mechanism should make classification benefit from its trial-specific weights. Our results show otherwise. As shown in Figure 1, we fix current-trial Values and every model parameter. Across eight decoders, cross-sample and training-mean replacement changes mean BA by only −0.51 to +0.93 pp. Some replacements even increase mean BA for ATCNet, EEG Conformer, CTNet, and TCFormer. These results directly show that current-trial weights fail to provide the classification benefit expected from an effective attention mechanism.
Classifier accuracy and mechanism validity are separate questions. The concrete cost of this failure is false mechanistic attribution: observed gains cannot be credited to trial-specific weighting, and the attention maps lack a decision-level basis. This can misdirect architecture design toward attention blocks or downstream compensators without verifying route use. Accuracy, generalization, and robustness evaluate the complete system; our audit determines whether the advertised mechanism contributes to its performance.
Cross-sample replacement is an in-distribution mechanism counterfactual. It bijectively reassigns valid held-out attention tensors, preserving their marginal distribution and route activity while breaking only their pairing with current-trial Values. A task-level response is required to claim functional benefit from those weights.
SOAMNet establishes route dependence without sacrificing task performance. Uniform attention supplies an analytic null, while cross-sample and training-mean replacement provide empirical tests that the architecture does not guarantee: a single path can still learn static weights, nearly uniform routes, or route-insensitive Values. SOAMNet passes both empirical tests on all four datasets and achieves the highest mean BA and κ among the compared methods. It therefore satisfies both requirements of the repair: the route is demonstrably used, and the classifier remains the strongest on average.
Response to Critical Weakness 2 and Question 2: global context and temporal selection
The global-context result supports the repair claim by revealing how the repaired route carries class evidence. Our repair criterion is functional dependence on sample-conditioned routing; local-peak discovery is a separate, narrower property. In SOAMNet, global context conditions scores and Values and reaches the logits only through the audited readout. Replacement therefore disrupts a genuine sample-specific coupling in the sole decision path.
The controls also separate global and temporal contributions. Independently trained global-only variants lose 0.95--2.45 BA points on every dataset, with all participant-paired intervals excluding zero; removing content Q--K scores costs 1.08--4.29 points, and evidence-matched swaps cost 1.32--3.12 points. Global context drives the principal route, and trial-specific temporal evidence adds a consistent classification gain.
The bypass experiment provides direct causal support for this interpretation. Adding an external evidence path collapses Rcore from 18.61--40.95 pp to approximately zero while reducing clean BA by only 0.68--2.97 pp. Route dependence disappears precisely when the classifier regains unrouted access to class evidence, as the proposed bypass mechanism predicts. The repair claim is therefore supported by both the controlled information path and the resulting classification utility.
Response to Question 3: clarification of the non-attention control
The matched non-attention head is a valid control because it matches SOAMNet in encoder, source evidence, capacity, and training protocol. The reported 3.51--5.22 pp gap is obtained with a same-encoder nonlinear full-trajectory pooling head. For each participant and seed, it uses the same encoder checkpoint and receives the complete feature and class-evidence trajectories used by SOAMNet. The control receives the source evidence sequence from which SOAMNet derives g and m; its freshly trained head applies a token MLP, temporal mean/max pooling, and a logit MLP.
The control uses identical splits, seeds, optimization caps, and validation-based checkpoint selection, and has slightly more active parameters than SOAMNet (135,632 versus 130,474 on BCIC-IV-2a). These matched conditions largely rule out a weaker classifier, fewer source features, lower capacity, or a smaller training budget as explanations for its lower BA across all four datasets. The comparison therefore supports added classification utility from the attention-based readout.