acceptodds
Under review as a conference paper at ICLR 2027

How Many Labelled Trials Is Euclidean Alignment Worth? A Within-Corpus Dose-Response for EEG and EMG Transfer

Abstract

Euclidean Alignment (EA) and its Riemannian relatives are near-universal preprocessing in EEG and EMG decoding, and are reported as accuracy gains over an unaligned cross-subject baseline. That baseline answers the wrong question. The question a practitioner faces is not does alignment help? but is borrowing donor data worth more than labelling a few trials from this user?, and answering it requires a column the alignment literature does not report: an in-domain k-shot decoder trained on the target subject alone. Across ten public configurations and 36 within-corpus cells, the crossing budget—the number of labelled target trials at which the in-domain decoder reaches the aligned cross-subject one—ranges from 12 to over 240 total trials and depends on the decoder, so no single constant answers the question and a paper reporting a gain without this column has not reported enough. We characterise what governs the budget. It is not class count: a between-corpus ordering by K looked compelling and was refuted by the withincorpus dose-response we ran to test it, which finds k * falling with K on high-SNR sEMG and flat on low-SNR EEG-MI. It is not modality either. What our data leaves standing is within-subject decodability, measurable in advance—though we show only that class count does not govern the budget, not that decodability does. Priced for the decoder a practitioner would actually deploy—tangent-space logistic regression on both arms, with the in-domain arm touching no donor data at all—the budget is 18–40 total labelled trials on every corpus tested, a 2.2× band across two modalities and K in 2, 4, 6. Read per user rather than per corpus, 60 total trials reaches the aligned level for 87% of sEMG and 70% of EEG-MI subject-cells, with 10% and 22% served by no tested budget—a ceiling set by our grid rather than an observed asymptote, and a lower bound on what the cross-subject route costs. All of this is interpretable only against a calibration of zero, which is the paper's second contribution: when the alignment reference and the fit are both in-domain, the gain is identically zero for any decoder invariant to the corresponding congruence—a two-line consequence of affine invariance, exact, falsifiable by one trial changing one label, and holding on 137 subject-fits (119 above a pre-declared informativeness floor) on two classifier families at every regularisation tested. A factorial then separates the two knobs the literature conflates: the reference governs exactness, the fit governs magnitude.

Primary area:
applications to neuroscience & cognitive science
Keywords:
domain adaptation, Riemannian geometry, affine invariance, EEG, EMG, transfer learning, brain-computer interfaces, negative results, reproducibility

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

This paper reports measurement-validity / evaluation for the manifold EEG/EMG cross-subject decoding pipelines. The ML contribution appears limited: most conclusions are established for a narrow family of manifold-based EEG/EMG pipelines. The paper is also very poorly written, with too many subtitles and sentences that are very hard to follow. I think this paper is not of interest to the ICLR machine learning community in EEG; it reads more like an engineering technique report.

Critical strengths

S1: Cross-subject decoding is an important research topic for EEG, and the results presented in this paper would be of interest to engineering researchers in EEG.

Critical weaknesses

W1: What is the paper's contribution related to the machine learning community? It does not propose a new alignment method or decoding method. Its main contribution is a simple measurement-validity / evaluation in the previous work. It claims to report quantities in the field; I did not see anything new from its perspective. The theory is also more or less unnecessary.

W2: Domain adaptation and foundation models are two main approaches in the EEG cross-subject decoding community. Domain adaptation has its own advantage of being small models and without the need for large computational resources. However, even in this setting, the comparison with classical EEG deep learning models and recent high-impact foundation models is also necessary.

W3: This paper is very poorly written, and I found it very hard to follow. There are subtitles like 'What this paper is not. '; 'The screen, and the two numbers it produces'; 'Reading this as a zero, not a result.'; 'Add the column. '.

Table2, Table3, Table 4, and Table5 violate the ICLR format requirements.

Questions for authors

Can you compare the alignment framework with classical EEG deep learning and foundation models? These models are also important for cross-subject decoding.

What is the paper's contribution related to the ICLR community?

Why do we need such analyses for large samples, the reasonable number of available data trials can provide a good estimation, in my opinion.

1 comment

Contribution. The paper proposes no method, by design: it supplies the in-domain baseline that published alignment gains omit (the domain-adaptation literature reported it a decade ago on ERP features, Wu 2017; the alignment literature never imported it — Related Work now says so), prices the budget at which own labels replace donors, and (new in this revision, Table 5) prices the pooled hybrid that the EA review itself recommends (Wu 2025). We accept that this is a measurement contribution and have removed language that implied more.

Deep and foundation models. Two additions. Appendix R reports a trained SPD network on BCI IV-2a: alignment changes a network with the identity basepoint by +0.114 and one with the per-domain basepoint by +0.0004, as the proposition predicts. Appendix S (Table 16) puts EEGNet on the budget axis, same splits as Tables 4–5: aligned EEGNet is +0.05 [+0.02, +0.08] above aligned TS+LR on binary IV-2a (+0.032 [−0.005, +0.071] on 4-class); EEGNet trained from scratch on the user's k trials is below the shallow own-trials decoder at every k; and a donor-pretrained EEGNet fine-tuned on the same trials keeps a +0.12–0.15 premium at the largest k, so for that decoder donors have not stopped paying at any tested budget. The budget claim is now scoped to the decoder family it was measured on. We have not run foundation models; the budget question for those is open, and we say so.

Format and writing. Tables 2–5 of the submission overflowed the text margin; all tables are now inside it. The main text's stand-alone qualification paragraphs are folded into the paragraphs they qualified (global comment, correction 6); the sentence-style paragraph headings remain, and a full rewrite is beyond the discussion window.

Why large-sample analysis. The question is not how many trials estimate a covariance but how many labelled trials from a new user make donor data unnecessary — Table 5 gives the answer (5–40 per class) and the regime where it fails (60-channel Weibo2014).

Reviewer_7tzaOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper looks into whether, for a new BCI or sEMG user, it is better to borrow "donor data" through Euclidean alignment or to label a few trials from that user. The question is interesting and practically relevant. However, the cost accounting is incomplete, several headline claims are not supported by the result tables, the scope is limited to classical pipelines, and the manuscript is very hard to read.

Critical strengths
  • Interesting question, indeed.
Critical weaknesses

Clarity and Presentation

  • Alright, this paper could be benefited from human writing.
  • There are many "leftover" from defensive wordings to a behind-scence reviewer (!) like e.g. "two earlier versions of this paper asserted opposite answers" (L316), and "the reviewer's framing" appears at L1131
  • Several captions do not match their figures. Figure 1 shows "0/105 fits changed" while the caption reports 0/137

Methodology

  • The scope is limited to classical covariance pipelines. the authors state their claims are "about alignment, not deep architectures". Given that deep and foundation models now dominate EEG decoding, I am concerned about how much the conclusions inform current and more popular practice.

Experiment and Results

  • Several mismatched part. The abstract states that class count "does not govern the budget", but the Table 5 caption reports that "the total at the crossing rises with K on both" corpora.
  • Weibo2014 is excluded as a "broken baseline". However, it is not well justified to me.

Minor Issues and Suggestions

  • Some numbers inconsistent across tables. For example, the binary IV-2a in-domain MDM arm reads 0.678 at k = 10 in Table 3 but 0.662 in Table 4.
Questions for authors

I don't have any questions that would expected to be answered given its current format, but would encourage the authors to revise the paper manually, linking with state-of-the-art, and with critical human thinking.

2 comments

Every factual point you raised was correct; thank you for checking the numbers.

Leftovers and captions. The references to earlier versions and the phrase addressed to an internal reader are removed; Figure 1 is regenerated (it was a stale six-corpus build; the caption's 0/137 was the correct count); the 0.678/0.662 pair came from two harnesses with different calibration draws — Table 4 is now recomputed on Table 3's test blocks, both captions say the residual differences are by draw, and the binary IV-2a nearest-mean crossing is 60 in both.

Abstract vs Table 5 (now Table 6). The abstract conflated the per-class budget k* with the total kK. Class count does not govern k; the total rises with K through its denominator. Both now say exactly that.

Weibo2014. You were right to doubt the exclusion. Re-run at α = 0 and 0.1: the in-domain nearest-mean arm is at chance at both, but the tangent-space in-domain arm is not (0.63 at k = 20 vs 0.68 aligned). Weibo2014 is now priced (Tables 4–5) as the configuration where donors still beat 20 own trials per class, and the headline is scoped accordingly.

Scope. We agree the study is of classical covariance pipelines and do not claim otherwise. Appendix R adds a trained SPD network, where the exact-zero prediction holds for the per-domain basepoint and fails for the identity basepoint; Appendix S adds EEGNet on the budget axis, where the budget conclusion is decoder-specific (the fine-tuning premium does not reach zero by k = 60) and the paper now scopes it to the tangent-space hybrid.

Reviewer_7tzaOpenReview ↗

Thanks

Thanks for the response. But I won't expect (and I think you will not as well) to raise the score, if reading the response feel like I am talking to AI, rather than a real person.

Reviewer_xKm3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper reexamines Euclidean Alignment (EA) for EEG and EMG transfer by comparing aligned cross-subject decoding with in-domain few-shot decoding. It establishes an invariance-based zero-effect result and empirically investigates the number of target-subject labels required to match aligned transfer performance.

Critical strengths
  1. The paper raises an important practical question: whether collecting a small number of target-subject labels is preferable to relying on cross-subject alignment.
  2. The analysis provides a clear mathematical explanation of alignment invariance and systematically investigates the effects of reference domains, classifier configurations, and regularization.
Critical weaknesses
  1. The central label-budget conclusion is supported by a much narrower evaluation than the overall study suggests. Although the paper investigates ten configurations, the highlighted 18–40-trial range for tangent-space logistic regression in Table 4 is established using only three evaluation settings drawn from two corpora, GRABMyo and BCI IV-2a. Several other datasets have censored crossing budgets or are excluded from pricing. Consequently, this relatively narrow range does not yet establish a broadly applicable relationship across different EEG/EMG tasks, recording conditions, and subject populations.

  2. The crossing-budget estimates are statistically unstable and lack sufficient uncertainty quantification. The crossing budget is obtained from discrete intersections of average accuracy curves rather than a statistically estimated threshold. Table 4 does not provide confidence intervals for the reported crossings, and some conclusions depend on accuracy differences of only a few thousandths. The supplementary analysis itself acknowledges substantial uncertainty and heterogeneous subject-level crossings. These limitations are particularly important because the crossing budget is presented as the paper's main practically actionable result.

  3. The proposed decision framework does not adequately represent realistic hybrid adaptation strategies.The main comparison treats aligned cross-subject decoding and target-only few-shot decoding as competing alternatives. However, practical systems can combine pretrained donor information, alignment, and a small number of labeled target trials. Such hybrid strategies may outperform either endpoint or change the estimated label budget substantially. Although the appendix includes a limited investigation of Riemannian Procrustes Analysis, the central dose-response analysis does not systematically consider donor-assisted few-shot adaptation or other stronger hybrid alternatives. Thus, the reported exchange rate is conditional on a restricted set of deployment choices.

  4. The theoretical result is largely a consequence of established geometric invariance rather than a new characterization of alignment effectiveness.The exact-zero result follows directly from congruence invariance of the affine-invariant Riemannian metric and equivariance of the Fréchet mean. The empirical verification is useful as a diagnostic, but it does not establish a new theoretical property of EA. More importantly, this exact result applies only to particular combinations of metric, reference, and decoder. The broader practical conclusions depend on empirical observations outside that invariant regime, where the current experiments do not yet provide a similarly comprehensive characterization.

Questions for authors

See weekness.

2 comments

Thank you — your three weaknesses are the ones we most needed to hear.

Narrow evidence for 18–40. Correct: three configurations of two corpora. The revision says "three of four configurations priced this way" and adds the fourth (Weibo2014), where the crossing is not reached at 40 total; the abstract no longer says "every corpus tested."

No uncertainty on crossings. Every k* now carries a 95 % subject-bootstrap interval (Tables 3–4, Appendix Q). The GRABMyo crossing is 3–5 trials per class; the EEG crossings are three- to six-fold wide, and the text now gives the width.

Hybrid strategies. We ran one: aligned donors pooled with the user's k trials in one tangent-space logistic regression, at equal weight and at target-mass fractions 0.1–0.5 (Table 5). The donor premium reaches zero at 5–20 (sEMG, 4-class EEG) or 40 (binary EEG) trials per class and is negative beyond. Weighting the user's trials more heavily raises k0 and never makes the premium positive past it. This is one family; the supervised Procrustes rotation is in Appendix P, and Appendix S prices the other way donors can enter a decoder — as the initialisation of a network that is then fine-tuned — where the premium does not reach zero on the grid (+0.12 at k = 60, binary IV-2a). Your point that the exchange rate is conditional on the deployment choices tested stands; the revision now shows one choice on which it changes sign and says so in §6.5.

The zero as a consequence of known invariance. Agreed; §6.1 now says it is not a result but the origin of the scale on which published gains are read. Appendix R now shows the same regime on a trained network with the TSMNet basepoint (ΔEA = +0.0004) against the identity basepoint (+0.114), which is the only new thing the zero lets us say.

Thank you for the reviewer's response, which addresses some of the issues. However, I still believe the quality of the paper falls short of the standard required for acceptance at ICLR. The entire manuscript requires revisions to its presentation to improve readability. I will maintain my rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author response and summary of revision

We thank the reviewers. All three found the paper hard to read, and Reviewers 7tza and mCsT identified concrete errors. We verified each against our source; every one was real, and every one is fixed in the revised PDF. The changes:

Corrections. (1) Figure 1 was a stale build of six configurations; it is regenerated and reads 0/137 across ten, matching its caption. (2) Table 4 is recomputed on the same test blocks as Table 3 with its own calibration draws and a grid extended to k = 60; the binary IV-2a nearest-mean crossing is k* = 60 in both (Table 4 previously read "never" because its grid stopped at 40), and both captions now state that the remaining differences between the two tables are by draw, not by protocol. (3) The abstract, §6.6 and the dose-response table (Table 5 of the submission, now Table 6) are reconciled: class count does not govern the per-class budget k*; the total k*K rises with K only through its denominator. (4) The sentences referring to earlier versions of the paper, and a phrase in Appendix P addressed to an internal reader, are removed. (5) Tables 2–5 of the submission overflowed the text margin; all tables are now inside it. (6) The main text's stand-alone qualification paragraphs ("What this paper is not", "How many of those 137 are informative", "Reading this as a zero, not a result", "What we do not claim") are folded into the paragraphs they qualified, as one or two sentences each; no number changed.

Uncertainty (xKm3). Every crossing budget in Tables 3 and 4 now carries a 95 % subject-bootstrap interval (Appendix Q): tight on GRABMyo (3–5 trials per class, n = 43), wide on the nine-subject EEG corpora (3–10; 10–60).

Hybrid arm (xKm3). New Table 5 prices donors + the user's own k labelled trials in one tangent-space decoder, on the same test blocks. The donor premium over own-labels-only is +0.12 to +0.23 at k = 1, reaches zero at k0 = 5–20 trials per class on GRABMyo and 4-class IV-2a and 40 on binary IV-2a, and is negative beyond: pooled donors cost up to 0.14 once a user has k0 own trials. The crossing budgets sit at or below k0, so the hybrid locates where donors stop paying rather than moving where own labels suffice.

Weibo2014 (7tza). Re-run at α = 0 and 0.1 with both decoders. The in-domain nearest-mean arm is at chance at both α; the tangent-space arm is not (0.63 at k = 20 vs 0.68 aligned). It is now priced in Tables 4–5 as the one configuration — 60 channels, pools of ≤ 40 trials per class — where donors still beat 20 own trials; the abstract says so.

Also added. The signal chain per corpus (Table 8, Appendix O). The exact-zero on a trained SPD network (Appendix R, Table 15): identity basepoint ΔEA = +0.114 (p = 0.004), per-domain basepoint +0.0004 (p = 0.69), on BCI IV-2a. A Related-Work paragraph placing the in-domain column against prior work: the ERP domain-adaptation literature did report target-only and pooled baselines (Wu 2017 — pooling without adaptation was generally worse than target-only, as in our Table 5, and adaptation matched at 5 labels what target-only reached at 100); the EA literature does not (He & Wu 2020; Junqueira et al. 2024; Wu 2025, whose calibration sweep keeps source data in every arm and recommends the pooled hybrid that Table 5 prices). It also credits Li et al. (2024, SPDIM) for the generative-model account of when alignment helps; our question is the narrower exact one.

A deep decoder on the budget axis (mCsT, 7tza). Appendix S (Table 16) runs EEGNet-8,2 on BCI IV-2a, binary and 4-class, on the same test blocks and calibration draws as Tables 4–5: trained from scratch on the user's k trials per class, and donor-pretrained then fine-tuned on the same trials. The result bounds our claim. The fine-tuning premium is +0.24 at k = 1 and still +0.12 [+0.06, +0.17] at k = 60 (binary) and +0.15 [+0.12, +0.18] at k = 40 (4-class): for this decoder there is no k0 on the grid, so "donors stop paying at k0" is a statement about the tangent-space hybrid, and §6.5 and the abstract now say so. The from-scratch network is below the own-trials TS+LR at every k (by 0.04–0.18), and the fine-tuned network beats the shallow hybrid by +0.07 to +0.10 at the largest k but is within +0.01 [−0.02, +0.04] of the shallow own-trials decoder on 4-class.

Not done. Foundation models, and hybrid families beyond pooling, fine-tuning and the Procrustes rotation of Appendix P. We agree this is a measurement paper, and the revision says so where the submitted version implied more.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.