acceptodds
Under review as a conference paper at ICLR 2027

Understanding and Controlling Batch Normalization Statistics for Source-Free Cross-Subject Motor-Imagery EEG Decoding

Abstract

Cross-subject motor-imagery electroencephalography (MI-EEG) decoders often lose accuracy on unseen users because intersubject distribution shifts affect both the input signals and the intermediate feature distributions. Although batch normalization (BN) is widely used in EEG decoders, the role of its running mean and variance in cross-subject generalization remains unclear. We isolate this role by decomposing a representative source-free test-time adaptation (TTA) procedure into target-statistics recalibration and affine-parameter optimization. Four strictly matched conditions and a source-statistics restoration intervention are evaluated on three datasets with three BN-based backbones. The target BN statistic recalibration alone reproduces nearly all of the composite adaptation gains at both the setting and matched-case levels. Disabling target statistic updates removes the mean gain, and restoring the source statistics largely eliminates the gain in all six settings for which restoration results are available. However, unconditional recalibration can still degrade individual target cases. We therefore treat BN recalibration as a selective deployment action and introduce a label-free risk gate calibrated by nested leave-one-subject-out (LOSO) evaluation. Using 23 target-time diagnostics, the gate selects either BN recalibration or source-only inference. It reduces exposure to negative transfer in all nine evaluated settings while retaining a setting-dependent proportion of the unconditional gain. The results identify target BN statistics as a measurable source of cross-subject adaptation gain and provide a practical means of controlling the associated risk.

Primary area:
applications to neuroscience & cognitive science
Keywords:
motor imagery EEG, cross-subject decoding, batch normalization, test-time adaptation, negative transfer

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4SdyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper looks int ot the role of source-free test-time adaptation in cross-subject motor imagery EEG decoding, focusing on batch normalisation (BN) statistics. The authors tried TENT (a standard tta method) in four levels: source-only, BN-only recalibration, TENT, and affine-only optimisation with fixed source statistics. It was studied on BCI IV 2a, BCI IV 2b, and OpenBMI, across three datasets EEGNet, ShallowConvNet, and ATCNet. The authors argued that BN-only recalibration closely matches TENT.

Although the authors have done some good experiments on BN, I am afraid the scope of this work is far less what would be expected from a ICLR paper in 2027. The main finding is largely anticipated by prior work, and the evidence is limited to BN-based CNNs on three small MI datasets.

Critical strengths
  • The four-condition design is clean. Four test conditions adhere to the same experimental settings in terms of the same checkpoint, split and seed.
Critical weaknesses

I am concerned about the limited novelty and scope of this paper, as well as the lack of genrealizability of the evidence.

Clarity and Presentation

  • There have been much space used for introducing BN or equivalent, which are particualrly not the key takeaways of this work.
  • The figures could be enhanced better, about the clarity, resolution, correct aspect ratio etc. I could see some AI traces associated with Figure 1 (but apologies if you did not apply AI).
  • There are many terms/phrases coming without context, and it adds to great difficulty to understand the whole ppaer (probably they come from AI's vocabulary, rather than human!), such as "pooled accucracy" in L77, apologies I may not be able to understand this term at first glance!

Methodology

  • Evaluating the testing adaptation of “old-fashioned” EEG deep models in 2026 would sounds a bit “old-fashioned” itself (especially considering foundation models have been dominating this area), and if not justfied or addressed well, it will sounds more like a report-like work rather than a paper bringing novel knowledge/insights to this community.
  • Further to the last, I believe this area (BN for EEG model adaptation) has been done a lot, and looking into the references in Related Work section, all before 2024, is there any more recent work on this? and what is the difference?
  • If TENT puls BN calibration is considered as one contribution, I am afraid people may not buy-in this. As I beleive this has been done significantly in test time adaptation based work (although may not in specific BCI EEG area)
  • Most recent EEG foundation models use LayerNorm rather than BN, so it is unclear how the takeaways transfer.

Experiment and Results

  • Just based on my experience of working these datasets, they are relatviely small, even different splits or small changes would result in big changes. After reading through this work, I am concerned that whether the takeaways from this paper can be as generalizable as possible to other datasets or tasks. So does it become clear (as said in L16 before this paper this issue remains unclear)?
  • Euclidean alignment is cited (L88) but not compared, although it is cheap, source-free and standard for cross-subject MI. Also, Tables 1 and 2 would be enhanced by adding the sd or confidence interval. I would encourage stronger baselines and statistics.

Minor Issues and Suggestions

  • Figure's readbility could be enhanced by increasing the font size.
  • It is difficult for me to understand the term "risk" in the context of this paper, and also several other terms. Does it mean probability, or literally "risk"??
Questions for authors

I would expect the authors

  1. Please clarify the scope of this work, its main takeaways and its novelty, in particular relative to Wimpff et al. (2024), the BN baseline in TENT, and Schneider et al. (2020).
  2. Does the near-equivalence of TENT and BN-only hold, if condutced under a sweep of learning rate,
Reviewer_5x8KOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper analyzes where the performance gain of source-free test-time adaptation (TTA) comes from in cross-subject MI-EEG decoding. The authors decompose TENT-style adaptation into BN statistics re-estimation and affine parameter optimization, and compare them across three datasets and three backbones. BN statistics re-estimation alone almost fully reproduces the TENT gain, whereas affine parameter optimization alone yields no gain. The authors also show that, despite positive mean gains, performance degrades in 16-38% of cases. To reduce this, they propose a gate based on 23 label-free diagnostics, which lowers the negative-transfer rate in all settings.

Critical strengths
  • It has been empirically known among EEG researchers that BN statistics re-estimation accounts for most of the TTA gain. This paper clearly demonstrates it through controlled experiments, which is its core contribution.

  • The case-level analysis exposes performance degradation in 16-38% of cases, which is meaningful from the perspective of real-world BCI deployment.

Critical weaknesses
  • The core finding is largely predictable from prior work such as AdaBN. The paper does not explain why recalibration degrades performance in some cases. The proposed gate does not improve re-estimation itself; it is a post-hoc selector that only decides whether to apply it. As a result, the paper offers no new perspective on fundamentally addressing inter-subject variability.

  • Affine parameter optimization is evaluated only under a single setting (learning rate 1e-4, 10 steps). The conclusion that it has no effect may therefore be an artifact of this setting. In addition, the gate is not compared with a random selection baseline matched in acceptance rate, so it is unclear how much of the NTR reduction comes from the gate’s judgment rather than from simply rejecting more cases. Furthermore, no reference such as supervised fine-tuning (as an upper bound) or full-parameter adaptation is provided, so the practical significance of the BN-only gain cannot be judged.

Questions for authors

Clarifications

  • Which optimizer was used for affine parameter optimization, and how did the authors choose the learning rate and number of steps? Does BN-only ≈ TENT still hold under stronger settings (e.g., larger learning rates or more steps)?

  • How are the six “legacy” fields among the gate diagnostics computed? Does the gate’s performance hold when they are excluded?

  • Do the results change if the gate diagnostics are computed only from the unlabelled adaptation subset, rather than also from the evaluation inputs?

Future work

  • How do the authors think inter-subject variability should be addressed more fundamentally, beyond deciding whether to apply recalibration?

  • It would be informative to compare BN re-estimation with supervised fine-tuning on labelled target data (as an upper bound) and with label-free full-parameter adaptation. This would show how much of the inter-subject gap BN re-estimation actually closes, and whether a broader update scope changes the conclusions.

Reviewer_91ZsOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines the role of batch-normalization (BN) running statistics in source-free, cross-subject motor-imagery EEG decoding. It decomposes a TENT-style test-time adaptation (TTA) procedure into (i) recalibration of BN running statistics and (ii) entropy-minimization of BN affine parameters. Under strictly matched conditions (same source checkpoint, adaptation/evaluation subsets, and seeds) across 3 datasets (BCI2a/2b, OpenBMI) x 3 BN-based backbones, BN-only recalibration reproduces the composite TENT gain. A negative control (affine optimization with source statistics frozen) yields no gain, and restoring source statistics after TENT eliminates the gain in the six BCI settings where the intervention was run. Because unconditional recalibration still harms individual cases (case-level negative-transfer rates, NTR, of 16.0-38.2%), the paper evaluates a nested-LOSO logistic gate over 23 hand-crafted, label-free diagnostics that routes each offline target case between BN recalibration and source-only inference; the gate reduces NTR in all nine settings (to 9.5-16.0%) while retaining 58.8-116.9% of the recalibration gain.

Critical strengths
  1. The matched-condition decomposition with a source-statistics restoration intervention is a rigorous and reusable audit protocol for attributing TTA gains.

  2. The case-level negative-transfer analysis exposes a deployment-relevant risk (NTR 16.0-38.2%) that pooled mean accuracy conceals; this is a useful warning for BCI practitioners.

  3. Gate fitting uses nested subject-level LOSO with no target labels, avoiding the most obvious leakage route for the risk-control experiment.

Critical weaknesses
  1. Limited methodological novelty and unclear distinction from existing work. That updating BN statistics improves performance under distribution shift is well established by adaptive BN, prediction-time BN adaptation, TTN, and online EEG TTA comparisons - all cited by the authors. The mechanistic conclusion is therefore a rigorous confirmation of a known phenomenon on EEG benchmarks, not a new principle. The risk gate is a standard logistic regression, and its optimization machinery is underspecified: the fold-level threshold-selection criterion is never stated anywhere in the paper; the contributions of the platform-dependent runtime feature and three legacy fields among the 23 diagnostics are not ablated; sensitivity to the 0.002 training-label threshold is admittedly unestablished. Moreover, the absolute gains are modest, consistent with prior systematic BN studies reporting real but small effects. Unless the authors can articulate a sharper delta over adaptive-BN/TTN-style work, the abstract and conclusion should reposition the contribution explicitly as a mechanism audit and risk characterization.

  2. The risk gate lacks sufficient validation against baseline methods. Simply keeping the full source model yields NTR = 0 and completely avoids negative transfer; the gate's value must therefore be demonstrated on the safety-utility frontier against trivial and simple baselines - always-source, always-BN, single-feature gates (e.g., a source-entropy threshold), and the archived fixed-threshold gate that is mentioned only as a reference. None of these comparisons is provided. The reported trade-offs are unfavorable in several settings: BCI2b/ShallowConvNet retains only 58.8% of the adaptation gain; BCI2a/EEGNet's NTR improvement is 16.0% -> 15.6% (almost certainly within noise given 225 correlated cases, at the cost of 20.4% of the gain); and no confidence intervals or significance tests are reported for any NTR change. The claim that the gate "reduces negative-transfer rates in all nine evaluated settings" is thus statistically unsupported at its weakest points.

  3. The transductive gate protocol limits deployment value. The diagnostics are computed on the locked, unlabelled evaluation inputs before the decision; such inputs are unavailable before deployment. The deployable alternative - computing the same diagnostics on the adaptation subset only - is untested, so the gate currently functions as a post-hoc audit tool rather than a deployable controller.

  4. The NTR statistic uses the wrong unit for the deployment question. Each case is a subject x source-seed x split-seed realization (25 correlated cases per subject in BCI2a/2b). Seed/split noise and user-level risk are conflated; subject-level NTR and gate performance are not reported, and the abstract's NTR figures can be misread as user-level risk.

Questions for authors
  1. What is the gate's performance after removing the runtime and three legacy diagnostic fields? How does it compare with reliability-aware TTA methods?

  2. Can the gate be driven by diagnostics computed on the adaptation subset only? What is the NTR/retained-gain cost relative to using the evaluation inputs?

  3. Please provide confidence intervals (at least subject-level bootstrap) for the NTR changes in Table 2, and report subject-level NTR with and without the gate. Is 16.0% -> 15.6% significant?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.