- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2(+1)
Read review
Paper summary
In this work, the authors propose TASER to address the problem of nonstationarity in cross-subject EEG deployment. In particular, the authors emphasize that the overlapping windows of the same trial are highly correlated; If adaptation is triggered once per window, an exception trial will repeatedly push the state update, and the result may depend on the extraction density of the window. TASER adapts a frozen EEGNet decoder through a shared latent translation of fixed class/phase prototypes. After recording predictions for an acquisition unit, it pools the embeddings into a shift observation and performs a recursive update with a bounded gain, with the gain informed by within-unit dispersion and window correlation. The evaluation covers 13 datasets, 153 retained participants, and four source-model seeds. Dataset-balanced accuracy is 70.00%, compared with 68.61% for NEO.
Critical strengths
- The paper connects EEG acquisition structure, causal predictions, and label-free adaptation in a compact design. I think this makes the work relevant to building BCI decoders that can operate continuously with modest overhead.
- The study spans MI, ERP/P300, and SSVEP across 13 datasets, using shared source models and natural stream order. This breadth provides a useful empirical foundation. I also appreciate the discussion of task-dependent behavior, small component gains, and the limits of offline replay, which helps readers understand the scope of the results.
Critical weaknesses
- I think the additional reliability attributed to trial-level updates is unclear currently. I am not yet convinced what additional statistical or decision-making advantage a completed trial provides over smoothing correlated windows with controlled cumulative influence. A completed trial can also contain a coherent common offset: this changes the shift observation in Equation (9) without changing the centered residuals used in Equations (11)-(17). I therefore remain cautious about the reliability interpretation of the gain adjustment and smoothing.
- In my opinion, the shared prototype model rests on strong assumptions with limited principled justification. The prototypes are source class/phase embedding means, treated as transferable, fixed discriminative geometry whose target changes are captured by a common translation. This approximation may have limited capacity for class-specific changes, altered temporal structure, and more complex EEG shifts. The filtering rationale does not establish a theoretical connection between EEG changes and this shared geometry, or sufficiently explain its validity conditions. The reported results support the method's empirical usefulness, but I would find its scope more convincing with direct evidence that a shared shift captures the relevant target changes.
Questions for authors
- Does NEO-Causal include the published NEO-Continual EMA variant? How does TASER compare with K=1 plus fixed trial-level EMA under matched readouts and cumulative update strength?
- All reported results use EEGNet. Do authors have paired results for another encoder family, or should the conclusions be scoped to EEGNet?
- For ERP, is there one embedding per event, and how are K and the class-phase prototypes defined? This also affects the interpretation of the approximately 0.33-point gain over Proto-frozen.
6 comments
Response to CW2 and QfA1
- Evidence for the shared-shift assumption (CW2)
To examine the shared-shift assumption directly, we measured each class’s displacement from its source prototype to its mean target embedding, using target labels for this diagnostic analysis. We quantified the fraction of displacement energy explained by a common translation, computed separately within each participant’s encoder space and then averaged with equal weight across datasets.
Across all datasets, the common component accounts for 75% of displacement energy in MI (95% CI [67%, 82%]) and 43% in SSVEP [27%, 58%]. The corresponding estimates in the class-discriminative subspace are 61% [52%, 70%] and 36% [23%, 50%]. In other words, the class displacements share a substantial common component, especially in MI, where a single translation explains three-quarters of their displacement energy. This is the component that TASER aims to correct. Related longitudinal MI-BCI findings show changes affecting both classes similarly (Tortora et al., 2022), with comparable temporal evolution across MI tasks and coherent trajectories in Riemannian space (Palatella et al., 2026). These observations support the plausibility of shared changes across classes, although they concern a different representation and experimental setting.
We tested whether a common translation estimated from one part of a target session improves classification on unseen trials from the same participant. Using labels for this diagnostic analysis, we estimated common and class-specific shifts on the first half of each session and evaluated both on the second half, keeping the encoder fixed.
Across all participants, the common shift improves balanced accuracy by 9.0 points on MI (95% CI [4.1, 14.5]) and 2.5 points on SSVEP [-2.2, 7.5], with an overall gain of 6.1 points [1.9, 10.4]. Class-specific corrections provide no conclusive additional benefit on MI (+2.1 points relative to the common shift, 95% CI [-0.9, 5.3]) and reduce SSVEP accuracy by 5.3 points [-10.1, -0.5].
In other words, per-class correction is inconclusive on MI and harmful on SSVEP, where many classes have few trials; a common shift needs fewer estimates, which suits TASER’s label-free causal setting.
All results are significant except the common shift's gain in SSVEP and the additional MI benefit of per-class correction (+2.1).
These findings support the predictive relevance of a shared translation, particularly for MI. We will include both analyses, with their limitations, and clarify the model's scope in the revision.
- NEO-Continual and the fixed-gain trial-level (QfA1)
We started from NEO as released by its authors, which reaches 67.86% on the nine MI/SSVEP datasets (natural stream). We then made it causal, updating the target mean after each decision so that it follows the same predict-then-update protocol as all other methods; this NEO-Causal version reaches 68.37% and is the one reported in the paper. We had considered only the officially released version, which was developed for computer vision. Following the reviewer's suggestion, we also implemented the published NEO-Continual EMA variant which reaches 68.09%. The difference from our NEO-Causal is not statistically significant (-0.28 points, 95% CI [, ]).
We also tested the requested simplified alternative, combining A1 and A3 (ablation): a single prototype per class () and one fixed-coefficient EMA update per trial. The coefficient is set to the mean TASER gain on source streams. This control retains the same trial-level evidence aggregation, while removing temporal segmentation and trial-dependent gain variation.
| Protocol | TASER | + fixed EMA | Difference |
|---|---|---|---|
| 1 s, 20 Hz | 69.5 | 69.0 | +0.47* |
| 0.5 s, 20 Hz | 66.8 | 66.5 | +0.37 |
| 1 s, 10 Hz | 70.5 | 69.4 | +1.12* |
Balanced accuracy (%), decisions at trial completion, with equal weight to each of the nine MI and SSVEP datasets. Differences are computed before rounding. 1 s/20 Hz. *: the 95% hierarchical bootstrap interval (dataset → participant → seed) excludes zero.
TASER exceeds this simplified alternative by 0.47 and 1.12 percentage points with 1,s windows at 20,Hz and 10,Hz, respectively; both hierarchical confidence intervals exclude zero. With 0.5,s windows, the difference is smaller (+0.37 points) and inconclusive. The combined ablation therefore supports a benefit from retaining temporal prototypes and adaptive gain under the two 1,s protocols, although it does not isolate which component accounts for that benefit.
Response to QfA2 and QfA3
- Generalization across encoders (QfA2)
We extended the evaluation to EEG Conformer, which combines convolutional feature extraction with temporal self-attention (Song et al., 2023), and two pretrained EEG foundation models: CBraMod, with criss-cross spatial-temporal attention (Wang et al., 2025), and LaBraM, which learns EEG representations through large-scale pretraining (Jiang et al., 2024). These choices test TASER beyond EEGNet across both architectural changes and pretrained representations. CBraMod and LaBraM were fine-tuned on source participants before target evaluation.
TASER achieves the highest dataset-averaged balanced accuracy among the four compared methods with all encoders. It ranks first on the MI and SSVEP paradigm averages for each encoder, whereas NEO performs better on ERP/P300 with CBraMod and LaBraM.
| Encoder | Task | Frozen | Proto | NEO | TASER |
|---|---|---|---|---|---|
| EEGNet | MI | 64.3 | 62.1 | 70.6 | 72.0 |
| ERP | 63.6 | 70.9 | 69.1 | 71.2 | |
| SSVEP | 64.7 | 60.8 | 65.5 | 66.3 | |
| Mean | 64.2 | 64.4 | 68.6 | 70.0 | |
| Conformer | MI | 63.1 | 62.5 | 68.5 | 70.2 |
| ERP | 65.0 | 69.3 | 69.0 | 69.6 | |
| SSVEP | 40.3 | 38.4 | 40.8 | 45.2 | |
| Mean | 56.7 | 57.2 | 60.1 | 62.3 | |
| CBraMod | MI | 60.1 | 59.9 | 64.3 | 65.9 |
| ERP | 65.8 | 65.9 | 69.1 | 66.0 | |
| SSVEP | 45.0 | 49.7 | 46.9 | 55.5 | |
| Mean | 57.2 | 58.6 | 60.4 | 62.7 | |
| LaBraM | MI | 63.2 | 62.9 | 68.2 | 69.8 |
| ERP | 64.5 | 65.0 | 67.7 | 65.4 | |
| SSVEP | 72.9 | 73.8 | 75.0 | 77.1 | |
| Mean | 66.8 | 67.2 | 70.3 | 70.7 |
Balanced accuracy (%), with datasets weighted equally. CBraMod and LaBraM: using pretrained models fine-tuned on source participants. Bold indicates the highest value in each row. LaBraM: 12 datasets (BNCI2014-002 excluded: no 10–20 channel labels).
- ERP/P300 embeddings, prototypes, and adaptation (QfA3)
Yes. For ERP/P300, each stimulus-locked event produces one embedding from its full 0.8,s epoch. Thus, and by construction, with one source prototype per class, target and non-target, and no temporal partition. Events are predicted individually; adaptation occurs only after all events in the acquisition group have been predicted, pooling their embeddings into one update. This differs from MI and SSVEP, where each trial contains overlapping windows and serves as both the prediction and adaptation unit.
Compared with Proto-frozen, TASER improves balanced accuracy by 9.90 percentage points on MI and 5.53 on SSVEP, whereas the updated ERP/P300 estimate is modest (+0.33 points; 95% CI [0.03, 0.86]). The ERP/P300 gain therefore reflects shared-shift adaptation, not temporal prototypes or overlapping-window aggregation.
To investigate the smaller ERP/P300 benefit, we estimated a participant-specific common shift using labels from the full target session and applied it to the source prototypes as a non-causal reference. This correction improves balanced accuracy by 8.8 points on MI, 4.9 on SSVEP, and only 0.6 on ERP/P300. These results show that correcting a common translation offers limited additional benefit on the evaluated ERP/P300 datasets, helping explain TASER’s modest gain over Proto-frozen.
We have decided to move the ERP/P300 experiments to the supplementary material and focus the main paper on MI and SSVEP. In our ERP/P300 implementation, each event provides a single embedding, so neither temporal segmentation nor the correction for overlapping-window redundancy is exercised. Prediction and adaptation also use different units: individual events are classified separately, whereas adaptation follows an acquisition group.
Response to CW1
We thank the reviewer for the careful assessment. The analyses prompted by CW1 and CW2 helped us better understand the effect of latent shifts and distinguish the behaviour of the update rule from the assumptions behind it. Below, CW refers to Critical Weaknesses and QfA to Questions for Authors.
In the original experiments, we made decisions at trial completion to keep the evaluation simple and avoid introducing a stopping rule and additional parameters. Following Reviewer bZBb's comments, we also evaluated decisions made before the trial ends, as relevant to interactive BCI control.
For this analysis, we introduced cumulative log-likelihood accumulation, following the principle of sequential evidence accumulation used in BCI decoding (Liu et al., 2017). For each class c, we sum the evidence from the windows observed so far: \begin{equation} S_j(c)=\sum_{i=1}^{j}\ell_i(c), \end{equation} where ℓi(c) is the log-likelihood score for class c from window i. A decision is issued when the gap between the two highest accumulated scores reaches a threshold θ; otherwise, the highest-scoring class is selected at trial completion.
We expanded the ablation to all nine MI and SSVEP datasets and all three temporal protocols. To examine whether trial-level updating offers an advantage over smoothing successive windows, we compared TASER with per-window updates whose gain bounds were rescaled as wf=1−(1−w)1/M, where M is the number of windows. This rescaling matches the cumulative retention associated with each bound over a complete trial.
We calibrate θ using source participants only, targeting mean stopping times of 25%, 50%, or 75% of the trial duration. The Full column uses all available windows. This protocol lets us compare trial-level and per-window adaptation under both early decisions and decisions at trial completion. Trial-level updating shows its clearest advantage when decisions are made early, particularly on MI. At trial completion, per-window updating with rescaled gains achieves similar accuracy.
Per-window updating with matched cumulative influence minus TASER (one update per completed trial), balanced accuracy (pp); 1 s/20 Hz; at 0.5 s and 10 Hz. *: 95% hierarchical bootstrap interval (dataset → participant → seed) excludes zero; †: only the within-dataset test of participant means gives p<0.05.
We also tested what happens when all windows within a trial are displaced together. After encoding the EEG, we added the same vector to every window embedding in selected trials. We did not change the raw EEG, labels, or trial order. We repeated this test across nine datasets.
First, we made the disturbance temporary. We shifted the embeddings for one trial, five consecutive trials, or a block covering 10% of the session. We then stopped adding the offset and supplied the original embeddings again. This experiment tests whether temporary disturbances leave TASER incorrectly adapted when normal data resume, and whether it can follow lasting changes in recording conditions.
For the larger tested offset, accuracy on those subsequent clean trials changed by +0.03, -1.05, and -1.83 percentage points, respectively. Thus, an isolated disturbance had little average effect, whereas repeated disturbances affected later predictions. After the 10% block, the distance between TASER’s shift estimate and its clean-run counterpart fell from 0.67 to 0.12 inter-class-distance units over 20 clean trials.
Second, we made the displacement permanent. We added the offset to every remaining trial, so successful adaptation required following the new position rather than returning to the original one. TASER tracked 90% of the imposed displacement within 26 trials. Its mean accuracy loss in the final half of the post-change period was 0.18 points. These results suggest that TASER could maintain decoding performance after a lasting change in recording conditions.
Without target labels or external feedback, TASER cannot verify whether an observed displacement should be corrected. Its uncertainty estimate reflects within-trial dispersion and does not detect a coherent offset. Instead, the bounded gain limits each update. Our tests show that temporary disturbances fade once clean data resume, while persistent changes can still be tracked, without labelled feedback. We will clarify this distinction in the manuscript.