acceptodds
Under review as a conference paper at ICLR 2027

TASER: Trial-Aware Shift Estimation and Re-Centering for Causal EEG Test-Time Adaptation in Brain-Computer Interfaces

Abstract

Electroencephalography (EEG) decoders are often adapted one window at a time. Yet overlapping windows are repeated views of the same trial, so one atypical trial can move the adapter many times. We introduce TASER, a causal and label-free method that adapts once per brain–computer interface (BCI) acquisition unit: a trial for motor imagery and SSVEP, or a stimulus group for ERP/P300. A frozen encoder maps the signal into a latent space, where source prototypes define the class geometry. TASER records the current prediction before adaptation, pools the associated embeddings into one observation, and estimates a shared shift between source and target representations. Window redundancy and within-trial dispersion determine the update strength. The shift then re-centers the prototypes for the next trial, without target labels, pseudo-labels, replay, or backpropagation. Across 13 datasets, 153 participants, and four encoder seeds, TASER reaches dataset-balanced accuracy, compared with for the closest baseline. It ranks first on eight datasets and second on three, with a median online latency of ms per decision. Updating after every window instead of once per trial reduces accuracy by points; removing the shift costs points. The ranking persists across three temporal resolutions. Therefore, these results indicate that respecting the BCI acquisition unit can make causal EEG adaptation both more accurate and less expensive.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG decoding, brain-computer interfaces, test-time adaptation, online adaptation, calibration-free BCI, causal evaluation, trial-aware adaptation, gradient-free adaptation, latent shift estimation, cross-subject transfer

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_Ak5jOpenReview ↗
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
  1. 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.
  2. 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
  1. 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.
  2. 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
  1. 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?
  2. All reported results use EEGNet. Do authors have paired results for another encoder family, or should the conclusions be scoped to EEGNet?
  3. 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
AuthorsOpenReview ↗

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.

  1. Trial-level updates and the interpretation of reliability (CW1)

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 cc, 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)\ell_i(c) is the log-likelihood score for class cc from window ii. A decision is issued when the gap between the two highest accumulated scores reaches a threshold θ\theta; 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/Mw_f=1-(1-w)^{1/M}, where MM is the number of windows. This rescaling matches the cumulative retention associated with each bound over a complete trial.

We calibrate θ\theta 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.

ProtocolTask25%50%75%Full
1 s, 20 HzMI-2.54*-1.01*-0.64*-0.29
SSVEP-0.71-0.22-0.34-0.37
0.5 s, 20 HzMI-1.88*-0.72-0.43-0.28
SSVEP-0.79-0.35-0.34-0.43†
1 s, 10 HzMI-1.19†-0.92†-0.56-0.32
SSVEP-0.76-0.25-0.34-0.16

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<0.05p<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.

AuthorsOpenReview ↗

Response to CW2 and QfA1

  1. 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.

  1. 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 [−1.56-1.56, 0.690.69]).

We also tested the requested simplified alternative, combining A1 and A3 (ablation): a single prototype per class (K=1K=1) 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.

ProtocolTASERK=1K=1 + fixed EMADifference
1 s, 20 Hz69.569.0+0.47*
0.5 s, 20 Hz66.866.5+0.37
1 s, 10 Hz70.569.4+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.

AuthorsOpenReview ↗

Response to QfA2 and QfA3

  1. 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.

EncoderTaskFrozenProtoNEOTASER
EEGNetMI64.362.170.672.0
ERP63.670.969.171.2
SSVEP64.760.865.566.3
Mean64.264.468.670.0
ConformerMI63.162.568.570.2
ERP65.069.369.069.6
SSVEP40.338.440.845.2
Mean56.757.260.162.3
CBraModMI60.159.964.365.9
ERP65.865.969.166.0
SSVEP45.049.746.955.5
Mean57.258.660.462.7
LaBraMMI63.262.968.269.8
ERP64.565.067.765.4
SSVEP72.973.875.077.1
Mean66.867.270.370.7

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).

  1. 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, Mt=1M_t=1 and K=1K=1 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.

Reviewer_bZBbOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TASER, a causal, label-free TTA method for cross-subject EEG BCI. The authors argue that adapting on every overlapping window counts the same trial many times, so adaptation should happen once per BCI acquisition unit. TASER builds time-segment class prototypes on frozen EEGNet embeddings and predicts a trial by averaging the distances of all its windows. After prediction, the difference between the trial mean and the source center is used as an observation of a shared shift. All prototypes are then translated by this shift, with a clipped gain set by window redundancy and within-trial dispersion. In LOSO pseudo-online evaluation on 13 MOABB datasets, TASER reaches 70.00% dataset-balanced accuracy, ahead of NEO. In the ablations, only per-window updating and removing the shift have a large effect.

Critical strengths
  • Ten methods are compared across three paradigms and 13 datasets under the same checkpoint, observation order, and predict-then-update protocol. This is a useful comparison framework from the perspective of real BCI deployment.
Critical weaknesses
  • The core concept and technical novelty do not hold up

    • In MI and SSVEP, trial length is fixed, so feeding the whole trial window as a single input would make existing TTA methods update once per trial automatically. The window-overlap problem comes from slicing windows at a 50 ms stride. Moreover, TASER only decides after the trial ends, so it does not use the main benefit of windowing, which is continuous output.
    • Prediction and evaluation are done per event, but updates are done per selection. If a single flash is regarded as a trial, this setting actually departs from the acquisition unit.
    • The time-segment prototypes and the redundancy-based gain have almost no effect, and the clipped gain is effectively a fixed-coefficient exponential moving average.
  • The experimental evidence for superiority is weak

    • On ERP, TASER and Proto-frozen differ by about 0.33p on average. Frozen's linear head may have been disadvantaged by class imbalance, and every baseline that shares this head may have been affected in the same way.
    • Baselines are not tuned equally, and some of them collapse. Standard alignment methods such as EA and the classical methods mentioned in the introduction are missing.
Questions for authors
  • What gain was used for per-window updating in A2? Was the same bound applied as for per-trial updating?
  • Please compare against an encoder that takes the whole trial as a single input combined with per-trial re-centering. If TASER is not clearly better than this, the core design of "train on short windows, adapt per trial" is not justified.
  • Do the results hold with encoders other than EEGNet?
11 comments
AuthorsOpenReview ↗

Response to CW1 and QfA2

We thank the reviewer for the careful assessment and the time invested in our work. Your comments helped us expand the study and identify where the manuscript needs clarification. We apologise for the length of our responses: we wanted to address each concern carefully and explain the evidence informing the planned revisions. We organise the rebuttal into four main points. Throughout, CW refers to Critical Weaknesses and QfA to Questions for Authors.

  1. Window-based decoding and early decisions (CW1 and QfA2)

In BCIs designed for interactive control, sliding windows provide successive estimates of the user's intention while EEG is still being acquired. These estimates can drive ongoing feedback or be combined before issuing a discrete command. Evidence accumulation has been used in BCIs to combine successive predictions and issue a decision when the accumulated evidence reaches a threshold. Windowing therefore serves an operational purpose: it allows the decoder to respond before the complete task interval is available.

Our use of overlapping windows follows the motivation of pseudo-online evaluation: processing recorded EEG incrementally to approximate the information available during online operation (Carrara and Papadopoulo, 2024).

We used all available windows for the primary trial-level evaluation to compare adaptation methods without introducing a separate stopping rule and its parameters. A whole-trial decoder remains valid for decisions made at trial completion.

To evaluate decisions before trial completion, we additionally introduced cumulative log-likelihood accumulation (Wald, 1945), (Liu et al., 2017), (Jiang et al., 2018). This rule adds the evidence from successive windows. For each window jj and class cc, we compute \begin{equation} S_j(c) = S_{j-1}(c) + \ell_j(c), \qquad S_0(c) = 0, \end{equation} where ℓj(c)\ell_j(c) is the log-likelihood score assigned to class cc by window jj. We issue a decision when the difference between the largest and second-largest accumulated scores exceeds a threshold θ\theta, selecting the class with the largest score. If the threshold is not reached, we select the highest-scoring class at trial completion.

We calibrate the stopping threshold θ\theta on source participants to target mean decision times of 25%, 50%, or 75% of the trial duration. A lower threshold generally permits earlier decisions, whereas a higher threshold requires more accumulated evidence.

ProtocolMethod25%50%75%Full
1 s, 20 HzTASER64.768.469.369.5
NEO63.867.267.968.4
Frozen61.363.964.564.5
Proto-frozen58.260.861.561.5
0.5 s, 20 HzTASER58.864.866.566.8
NEO56.061.362.963.4
Frozen53.858.058.858.9
Proto-frozen52.555.957.057.1
1 s, 10 HzTASER65.269.470.470.5
NEO63.667.068.168.3
Frozen60.663.363.863.8
Proto-frozen58.760.861.361.3

Balanced accuracy (%), equal weight to each of the nine MI/SSVEP datasets

TASER achieves the highest balanced accuracy at the earliest stopping setting under all three protocols, exceeding NEO by 0.9, 2.8, and 1.6 percentage points, respectively. These results support TASER's usefulness when decisions must be made from limited evidence.

We also ran the requested whole-trial comparison. EEGNet was retrained with the complete trial as a single input. The whole-trial encoder produces one embedding per trial, which is classified before the re-centering state is updated. Frozen uses the corresponding source-trained classifier without target adaptation. These results support the design of training on short windows and adapting once per trial.

Sliding Windows

ParadigmFrozenTASER
MI64.372.0
SSVEP64.766.3

Whole trial

ParadigmFrozenRe-centering
MI65.870.4
SSVEP61.959.9
AuthorsOpenReview ↗

Response to QfA1 and CW3

  1. Per-window gain and extended ablations (QfA1 and CW3).

Yes. In the original A2 ablation, per-window updates used the same gain bounds as per-trial updates. Applying these bounds repeatedly within a trial increases its cumulative influence. We therefore extended the analysis to test smaller per-window gains. The revised A2 reported below rescales the bounds as \begin{equation} w_f = 1-(1-w)^{1/M}, \end{equation} where ww is the corresponding per-trial bound and MM is the number of windows. This matches the cumulative retention of the previous state across MM window updates to that of one trial-level update.

Following the concern in CW1 about waiting for trial completion, we also extended the ablations to early decisions using the cumulative log-likelihood accumulation described above.

The expanded analysis covers all MI and SSVEP datasets, with both decisions at trial completion and decisions triggered by accumulated evidence. To assess whether the findings depend on our windowing configuration, we repeated the ablations under all three protocols: 1,s/20, Hz, 0.5,s/20, Hz, and 1,s/10, Hz. The table reports TASER's balanced accuracy and its difference from each ablated arm.

ProtocolArm25%50%75%Full
1 s, 20 HzA0 TASER acc. (%)64.768.469.369.5
A1 K=1K=1+0.01-0.15-0.26†-0.24†
A2 matched step-1.72*-0.66†-0.51*-0.32†
A3 fixed gain-0.20†-0.26†-0.20-0.22
A4 cumulative-0.72*-0.67*-0.57†-0.94*
A5 no adaptation-6.1*-6.6*-6.9*-6.9*
0.5 s, 20 HzA0 TASER acc. (%)58.864.866.566.8
A1 K=1K=1-0.33-0.41†-0.37†-0.30
A2 matched step-1.40*-0.56-0.39-0.35
A3 fixed gain-0.02-0.07-0.12-0.09
A4 cumulative-0.46-0.24-0.36-0.52
A5 no adaptation-5.9*-6.9*-7.2*-7.3*
1 s, 10 HzA0 TASER acc. (%)65.269.470.470.5
A1 K=1K=1-0.98†-0.88†-1.01†-0.85†
A2 matched step-1.00†-0.62†-0.46†-0.25
A3 fixed gain-0.13-0.22†-0.30†-0.29
A4 cumulative+0.07-0.14-0.26-0.46
A5 no adaptation-5.7*-6.5*-6.8*-6.9*

A0: balanced accuracy (%). A1–A5: ablated arm minus TASER (pp); negative values indicate lower accuracy. Tested over MI/SSVEP. *: 95% hierarchical bootstrap interval (dataset → participant → seed) excludes zero; †: only the uncorrected within-dataset test on participant means gives p < 0.05.

Alongside A2-matched, we tested a per-window step selected on source participants, choosing the best-performing setting among the candidates.

TASER exceeds both alternatives by 1.55–1.72 points at 25% and 0.50–0.79 points at 50–75%, mostly with hierarchical intervals excluding zero. At trial completion (1 s/20 Hz), the advantage shrinks to 0.32–0.34 points (participant-level p < 0.05; hierarchical intervals include zero) but keeps its sign across protocols and paradigms.

Per-window variantMedian gain25%50%75%Full
A2-matched0.0020-1.72*-0.66†-0.51*-0.32†
A2, step selected on source–-1.55*-0.79*-0.50*-0.34†
TASER accuracy (%)64.768.469.369.5

Per-window variant minus TASER variant (balanced accuracy, pp); 1 s windows at 20 Hz. Tested over MI/SSVEP. *: 95% hierarchical bootstrap interval (dataset → participant → seed) excludes zero; †: only the within-dataset test of participant means gives p<0.05p<0.05.

AuthorsOpenReview ↗

Response to CW3-1, CW3-2 and CW3-3

Temporal prototypes (CW3-1)

The expanded A1 ablation shows that the contribution of temporal prototypes differs between paradigms. At trial completion, their effect on MI is negligible across all three protocols. On SSVEP, the corresponding gains are +0.50, +0.81, and +1.92 percentage points for 1,s/20, Hz, 0.5,s/20, Hz, and 1,s/10, Hz, respectively. All three reach p<0.05p<0.05 in the uncorrected participant-level tests, although their hierarchical bootstrap intervals include zero. Under 1,s/10, Hz, the SSVEP gains also appear at every early-stopping setting (+1.87 to +2.22 points). These results suggest that temporal prototypes are more useful for SSVEP than MI. Stimulus-locked SSVEP responses contain more reproducible temporal patterns that segment-specific prototypes can exploit, whereas variable MI timing makes this subdivision less useful. We will report this task-specific effect.

ProtocolTask25%50%75%Full
1 s, 20 HzMI-0.16+0.02-0.01-0.03
SSVEP+0.24-0.35-0.57†-0.50†
0.5 s, 20 HzMI-0.50-0.35†+0.02+0.11
SSVEP-0.12-0.49-0.87†-0.81†
1 s, 10 HzMI-0.11-0.08-0.04-0.00
SSVEP-2.06†-1.87†-2.22†-1.92†

A1 variant minus TASER (K=1K=1), balanced accuracy (pp)

Adaptive gain (CW3-2)

The expanded A3 ablation compares TASER's adaptive gain with a fixed gain matched to its source-side mean. With 0.5,s windows, the differences are small, and none reach significance under either test. With 1,s windows, the clearest benefits occur on MI at intermediate stopping settings: +0.44 points at the 50% setting under 20, Hz, and +0.40 and +0.50 points at the 50% and 75% settings under 10, Hz. The +0.40-point gain has a hierarchical bootstrap interval excluding zero; the other two reach significance only in the uncorrected tests. SSVEP differences remain small and inconclusive. The adaptive gain therefore provides a modest, setting-dependent benefit, with weaker evidence than the contribution of adaptation itself.

ProtocolTask25%50%75%Full
1 s, 20 HzMI-0.30†-0.44†-0.32-0.15
SSVEP-0.08-0.03-0.05-0.30
0.5 s, 20 HzMI-0.02-0.08-0.27-0.03
SSVEP-0.02-0.06+0.07-0.18
1 s, 10 HzMI-0.19-0.40*-0.50†-0.38
SSVEP-0.05+0.01-0.04-0.18

A3 minus TASER (gain fixed to the mean TASER gain on the source streams), balanced accuracy (pp)

Gain variability after clipping (CW3-3)

We examined the applied gain after clipping, trial by trial. Its behaviour differs across paradigms. In MI, no fold has a constant gain: approximately 31% of updates reach the upper bound, 42% reach the lower bound, and 26% lie between them. The gain is also negatively associated with within-trial dispersion (ρ=−0.56\rho = -0.56). Thus, clipping does not reduce the MI update to a fixed-coefficient EMA. The reviewer’s characterisation applies better to SSVEP, where 75% of updates reach the upper bound, and to ERP, where the gain is approximately constant. Then, in MI, variations in within-trial dispersion can produce a more variable gain, whereas in SSVEP the upper bound more often limits the update. We will explicitly report this distinction.

Reviewer_z8mkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes TASER, which is a test-time adaptation method for BCI. In contrast to existing methods, TASER treats a trial as one unit and adapts only once instead of window or segment wise adaptations.

Critical strengths
  • Rather than treating overlapping windows as independent observations and updating after each one, the method aligns adaptation with the acquisition unit of the BCI: one label-free update per trial (or per stimulus group for ERP/P300), made only after the trial’s prediction is recorded. This formulation reduces the need for continuous segment-wise adaptation and makes the proposed approach more practical for real-world deployment.

  • Additionally, the proposed approach does not need backpropagation, pseudo-labels, or a replay buffer. Its only evolving state is one latent shift vector and one scalar. Due to that, the approach seems computationally efficient and practical compared to other TTA or adaptation approaches.

Critical weaknesses
  • Although TASER is best on four of the five motor imagery datasets, its advantages are less clear on the other paradigms. Additionally, the reported performance gains appear relatively modest compared to the large standard deviations, raising questions about their statistical significance and consistent gains across different tasks and datasets.

  • The clarity of the presentation could be substantially improved. Figure 1 is a pipeline/flow diagram of processing stages and does not convey the underlying intuition. A geometric illustration would make the method much easier to grasp, for example a 2D latent space showing source prototypes, a shifted target cluster, and the prototypes and decision boundary translating over successive trials.

Questions for authors
  • In the cases where TASER outperforms other baselines, are the gains statistically significant?

  • All experiments are conducted using a single EEGNet encoder. Does the proposed approach generalize across different EEG encoders, model architectures, and EEG FMs?

  • Could the authors visualize the latent space (e.g., a PCA projection for a representative participant) showing the source prototypes, the target embeddings, and how the shifted prototypes and decision boundary evolve over the session compared with Frozen? How do the shift trajectory and accuracy vary?

3 comments
AuthorsOpenReview ↗

Response to CW1 and QfA1

We thank the reviewer for the constructive suggestions. We particularly appreciated the recommendation to complement the pipeline diagram with a geometric illustration, which helped us explain TASER's mechanism more clearly. Below, we address the concerns and describe the additional analyses and planned revisions. We use CW for Critical Weaknesses and QfA for Questions for Authors.

  1. Statistical significance and consistency of the gains (CW1 and QfA1)

TASER achieves higher mean balanced accuracy in all 90 comparisons, with 87 reaching significance in the Holm-adjusted participant-level tests. The only exceptions concern NEO, Frozen, and RoTTA on SSVEP under the primary protocol (1,s/20,Hz). Under both alternative protocols, every reported comparison reaches adjusted p<0.05p<0.05. Hierarchical 95% confidence intervals also lie above zero in 80 comparisons; the other seven significant tests have intervals that include zero. Against NEO, TASER improves MI accuracy in four of five datasets under each protocol, with positive differences for 71--79% of participants.

BCI performance is strongly participant-dependent: some users achieve high decoding accuracy, while others obtain substantially lower performance (Blankertz et al., 2010; Vidaurre and Blankertz, 2010). The large standard deviations in our results are consistent with this variability. We will report the complete statistical analysis in the supplementary material.

1 s / 20 Hz

BaselineMIERPSSVEPAll
NEO1.3†2.1†0.81.4†
Frozen7.7*7.7*1.65.8*
Proto-Frozen9.9*0.3*5.5*5.6*
T3A10.9*1.6*15.8*9.6*
RoTTA7.5*9.0*1.86.2*
T-TIME3.7*8.0*11.5*7.4*
BFT-A5.6*10.5*7.8*7.8*
BFT-D5.7*9.4*7.3†7.3*
PLTA4.7*1.4†15.7*7.1*

0.5 s / 20 Hz

BaselineMISSVEPAll
NEO2.2†5.0*3.4*
Frozen8.7*7.0*8.0*
Proto-Frozen10.2*9.2*9.8*
T3A9.4*14.0*11.4*
RoTTA9.0*6.8*8.0*
T-TIME5.5*17.6*10.9*
BFT-A5.1*11.7*8.0*
BFT-D5.3*11.0*7.8*
PLTA5.2*20.6*12.0*

1 s / 10 Hz

BaselineMISSVEPAll
NEO1.8†2.5*2.1*
Frozen8.4*4.4*6.6*
Proto-Frozen10.0*8.1*9.1*
T3A10.9*18.5*14.2*
RoTTA7.2*3.5*5.5*
T-TIME4.4*12.5*8.0*
BFT-A6.0*9.1*7.4*
BFT-D6.2*8.8*7.4*
PLTA5.4*18.9*11.4*

TASER −- baseline (pp). Paired tests: Holm correction over nine baselines per column; CIs: hierarchical bootstrap. ∗^\ast: pHolm<0.05p_{\rm Holm}<0.05, CI above zero; †^\dagger: pHolm<0.05p_{\rm Holm}<0.05, CI includes zero; unmarked: not significant. All: equal-weight mean across datasets included in each protocol.

AuthorsOpenReview ↗

Response to CW2 and QfA3

  1. Geometric interpretation and latent-space visualisations (CW2 and QfA3)

We will add a 2D geometric illustration alongside Figure 1, showing source prototypes, shifted target clusters, and the decision boundary before adaptation, after several trials, and later in the session. It will illustrate how all prototypes move by the same estimated translation, preserving their relative distances while the boundary translates without rotating. The encoder remains frozen, and each update follows the current prediction and applies from the next trial onward.

The schematic will use illustrative data. To complement it, we will include PCA projections of real EEG embeddings for each of the nine MI and SSVEP datasets. Each projection will show a representative participant selected by a fixed rule: the participant with the median TASER-minus-Frozen gain in that dataset. The plots will display target embeddings, source prototypes, and translated prototypes at successive points in the session.

To show how the shift trajectory and accuracy evolve over the session, we divided each target session into four equal quarters (Q1–Q4) and computed, within each quarter, the mean norm of TASER's shift, divided by the mean distance between source class prototypes, and the balanced accuracy of TASER and Frozen. The shift is estimated within the first quarter and then remains stable. Frozen's accuracy is flat across the session, whereas TASER exceeds it in every quarter. The gain is smallest in Q1, while the shift is still being estimated, and is consistent on MI (5.6–8.8 pp) but smaller and more variable on SSVEP. These results show that TASER establishes much of its correction early in the session and maintains an advantage over Frozen thereafter, with more consistent benefits on MI than SSVEP.

Q1Q2Q3Q4
∥δt∥\Vert\delta_t\Vert / inter-class distanceMI0.851.051.021.04
SSVEP0.350.610.640.61
TASER −- Frozen (pp)MI5.68.88.67.9
SSVEP0.63.00.62.6
AuthorsOpenReview ↗

Response to QfA2

  1. Generalisation 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.

EncoderTaskFrozenProtoNEOTASER
EEGNetMI64.362.170.672.0
ERP63.670.969.171.2
SSVEP64.760.865.566.3
Mean64.264.468.670.0
ConformerMI63.162.568.570.2
ERP65.069.369.069.6
SSVEP40.338.440.845.2
Mean56.757.260.162.3
CBraModMI60.159.964.365.9
ERP65.865.969.166.0
SSVEP45.049.746.955.5
Mean57.258.660.462.7
LaBraMMI63.262.968.269.8
ERP64.565.067.765.4
SSVEP72.973.875.077.1
Mean66.867.270.370.7

Balanced accuracy (%), mean over datasets. CBraMod and LaBraM (pretrained foundation models, fine-tuned on the sources). LaBraM: 12 datasets (BNCI2014-002 excluded: no 10–20 channel labels).

We will include these additional backbone comparisons and their evaluation settings in the supplementary material.

References

Blankertz, B., Sannelli, C., Halder, S., Hammer, E. M., Kübler, A., Müller, K.-R., Curio, G., & Dickhaus, T. (2010). Neurophysiological predictor of SMR-based BCI performance. NeuroImage, 51(4), 1303–1309.

Jiang, W.-B., Zhao, L.-M., & Lu, B.-L. (2024). Large brain model for learning generic representations with tremendous EEG data in BCI. In International Conference on Learning Representations (ICLR).

Song, Y., Zheng, Q., Liu, B., & Gao, X. (2023). EEG Conformer: Convolutional transformer for EEG decoding and visualization. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 31, 710–719.

Vidaurre, C., & Blankertz, B. (2010). Towards a cure for BCI illiteracy. Brain Topography, 23(2), 194–198.

Wang, J., Zhao, S., Luo, Z., Zhou, Y., Jiang, H., Li, S., Li, T., & Pan, G. (2025). CBraMod: A criss-cross brain foundation model for EEG decoding. In International Conference on Learning Representations (ICLR).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Experimental Evidence

We thank all reviewers for their careful reviews and constructive feedback, which helped us improve the manuscript and strengthen the evaluation. We apologize for the lengthy responses and delays in some updates. We have made every effort to address all concerns thoroughly and explain the additional analyses and revisions as clearly as possible.

In response to Reviewer z8mk, we added the statistical analysis, including paired tests with multiple-comparison correction and hierarchical confidence intervals. Against NEO (the strongest overall baseline), TASER improves dataset-balanced accuracy by 1.1, 3.4, and 2.1 percentage points across the three temporal protocols, with significant gains under the latter two. We also updated the main figures and added a geometric interpretation, real-data latent-space visualizations, and analyses of how the shift and accuracy evolve throughout a session.

In response to Reviewer bZBb, we expanded the evaluation to early decisions targeting 25%, 50%, and 75% of trial duration under all three temporal protocols, with TASER achieving the highest dataset-averaged balanced accuracy among the compared methods at each setting. We also conducted extended ablations to isolate the contribution of each component. The extended ablations show complementary benefits from TASER’s components: temporal prototypes improve full-trial SSVEP accuracy by 0.50–1.92 points, trial-level updating adds 1.00–1.72 points at the earliest stopping setting, and recent-weighted tracking adds 0.46–0.94 points at trial completion. In particular, removing shared-shift adaptation reduces balanced accuracy by 5.7–7.3 percentage points. Additional simulated-artifact experiments show that adaptive gain reduces contamination of the adaptation state by 11–37% relative to fixed-gain updates. We strengthened the baseline comparisons, added PMean and MPM-LDA for MI, and evaluated Euclidean Alignment. TASER outperforms both classical MI baselines on average, while EA + TASER achieves the highest mean balanced accuracy in each evaluated paradigm.

In response to Reviewer Ak5j, we directly examined the shared-shift assumption. Label-informed diagnostic analyses show that a common translation explains approximately 75% of displacement energy in MI and 43% in SSVEP. Estimating this correction on the first half of a session and evaluating it on the second improves balanced accuracy by 9.0 points on MI; the 2.5-point improvement on SSVEP remains inconclusive. These analyses support the approximation, particularly for MI. We implemented NEO-Continual, finding no significant difference from causal NEO, and tested one prototype per class with fixed-gain trial-level updates. TASER outperforms this simplified alternative, with statistically significant gains under two of the three protocols.

Finally, we reassessed ERP/P300. Class imbalance in the original prediction head affected the comparison with Frozen, while the adaptation-specific gain over Proto-Frozen is only 0.33 percentage points. Even a label-informed, full-session common-shift correction improves ERP/P300 accuracy by only 0.6 points, compared with 8.8 on MI and 4.9 on SSVEP. This indicates limited benefit from correcting a shared translation on these ERP/P300 datasets. Moreover, the ERP/P300 implementation uses a single embedding per event and does not exercise temporal prototypes or overlapping-window aggregation. We therefore moved these experiments to the supplementary material and focused the main paper on MI and SSVEP.

Addressing all three reviewers’ concerns about encoder generalization, we extended the evaluation to EEG Conformer, CBraMod, and LaBraM. TASER achieves the highest dataset-averaged balanced accuracy among the four compared methods: 62.3% versus 60.1% for NEO with Conformer, 62.7% versus 60.4% with CBraMod, and 70.7% versus 70.3% with LaBraM. It leads the MI and SSVEP averages with every encoder, while NEO performs better on ERP/P300 with CBraMod and LaBraM. These results demonstrate benefits beyond EEGNet.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.