acceptodds
Under review as a conference paper at ICLR 2027

Escaping the Capacity Ceiling: Routing on the Stiefel Manifold for Bilinear SPD Layers

Abstract

Deep networks on the symmetric positive-definite (SPD) manifold promise expressive representations by encoding data geometry as an inductive bias, but stacking BiMap layers with the standard ReEig nonlinearity often adds no capacity: on real, preconditioned EEG data, ReEig rarely activates, so the stack behaves as a single layer at any depth. In the worst case, when domains share no discriminative directions, we prove a single filter has a capacity ceiling, so it cannot fully align every domain at once. To overcome that, we propose SCAP (Stiefel Cross-Attention Pool), a layer implementing a family of Stiefel filters by combining a pool of experts into a sample-specific bilinear map via cross-attention. We show that it matches a per-domain filter bank to first order with fewer experts than domains when domain-optimal filters span few directions near a shared tangent-space basepoint; in the worst case, its alignment empirically stays nearly flat as domains grow, escaping the fixed-filter ceiling. Naively trained, however, this routing can collapse to a fixed filter; we diagnose why and adapt three mechanisms to mitigate it. SCAP significantly improves balanced accuracy over fixed-filter SPDNet on all five cross-domain EEG motor-imagery datasets, and matches or exceeds three domain-adaptive baselines on four out of five.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
SPD deep networks, Bilinear map, Stiefel manifold, Cross-attention

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_srQUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies the problem with SPDNet-based networks that are used for EEG decoding. It builds upon the assertion that layers (BiMap->ReEig) within an SPDNet often act as a single BiMap filter and therefore SPDNet performance does not improve by adding more layers. Starting from this point, the paper shows that a single BiMap has a capacity ceiling. Specifically, in the case of a single BiMap filter \in St(n,k) that is shared across m domains, the average per-domain subspace overlap is at most k/m (conditioned on mutually orthogonal subspaces, the worst-case). The paper next proposes a scheme to break through the capacity ceiling where instead of using a single BiMap filter, the input is routed through K Stiefel-manifold experts. These experts interact with each other via cross-attention and are combined as a Riemannian barycenter at learned anchor. The paper also address three routing-collapse where the K experts do not specialize. In this case the K attention weights become roughly uniform and set of experts is essentially reduced to a single filter. The proposed scheme is evaluated on five EEG datasets and the proposed scheme outperforms SPDNet.

Critical strengths
  • The paper makes an important observation that stacking BiMap->ReEig in SPDNet is reduced to a single BiMap filter in practice because ReEig never engages.
  • Lemma 1 that shows the capacity ceiling of using a single filter is easy to understand and is useful for practitioners.
  • Section 6 that discusses routing collapse failure and the three proposals for addressing this routing collapse are important.
Critical weaknesses
  • Experiments are conducted on motor-imagery EEG. The paper, however, claims that the method is equally useful for medical imaging and radar applications. No results are presented to support this claim.
  • MUSUME (Yamamoto et al., 2021 ) also uses multiple Stiefel filters for EEG. This should be included in the comparison table.
  • The paper will benefit from an analysis of what each of the K experts learns.
  • How would the claim about ceiling capacity hold when the subspaces are not mutually orthogonal?
Questions for authors
  • On real EEG, how close to mutually orthogonal are the per-domain optimal subspaces U_d?
  • Appendix H mentions that domain identity is the dominant routing signal. Is there a way to confirm if the proposed method is learning a per-domain filter bank? If so, how would this method compare against simple per-domain BiMap, each of which is trained for a single domain.
  • How sensitive is this method to the choice of K?
2 comments
AuthorsOpenReview ↗

Response with pointers to revised article pdf

W1 (medical imaging and radar claims). We agree that our results cover motor-imagery EEG only. The Discussion mentioned medical imaging and radar as settings where the same depth collapse could plausibly occur, as a direction for future work. To avoid any overreach, the revised Conclusion removes this conjecture and keeps these domains only in the list of future directions.

W2 (MUSUME as baseline). We implemented MUSUME from the paper and ran it on the same folds and inputs as every method of Table 1, as one model shared by all domains. MUSUME learns several Stiefel filters, fits an MDM classifier in each of their subspaces, and keeps the filter with the best validation accuracy. At test time it applies this single filter to every trial, while SCAP builds a filter for each trial. We use the published number of filters (4) and the subspace dimension of SPDNet and SCAP.

WeiboGW09BNCI15BNCI14Schirr.
SPDNet79.668.476.180.086.2
MUSUME76.859.473.473.567.2
input-routed SCAP (proposed)83.374.780.285.293.5
SCAP − MUSUME (Holm p)+6.5 (.061)+15.3 (.010)+6.8 (.010)+11.7 (.001)+26.3 (< .001)
time per run (s), MUSUME17351867781185
time per run (s), SCAP21661812135

MUSUME scores below SPDNet and SCAP on all five datasets. SCAP exceeds it by 6.5 to 26.3 points, significantly on four datasets. Giving MUSUME as many filters as SCAP has experts changes its accuracy by at most 0.3 points. MUSUME also takes 4 to 9 times longer than training SCAP, mostly to optimise its cost over pairs of neighbouring trials. MUSUME is now in Table 1 of the revised PDF (details in Appendix I).

W3 (what each expert learns). We analysed how the experts are used, on the runs of Table R1. Across trials, every expert is in use. The effective number of experts, the exponential of the entropy of the routing weights averaged over all trials, is 7.9 of 8, 7.9 of 9, 14.4 of 16, 7.8 of 8 and 10.0 of 12 (Weibo / GW09 / BNCI15 / BNCI14 / Schirr.). Within a single trial, the weight concentrates on fewer experts (Table R1, row H̄), and each domain has a preferred blend (row r_d). The experts therefore act as building blocks shared across domains. Appendix H now reports these quantities per dataset, with their discussion.

AuthorsOpenReview ↗

Response with pointers to revised article pdf ,2

W4 / Q1 (non-orthogonal subspaces; how close to orthogonal on real EEG).

Lemma 1 extends to overlapping subspaces. Let P=∑dUdUd⊤P = \sum_d U_d U_d^\top. Since sd(W)=tr⁡(W⊤UdUd⊤W)s_d(W) = \operatorname{tr}(W^\top U_d U_d^\top W), we have ∑dsd(W)=tr⁡(W⊤PW)\sum_d s_d(W) = \operatorname{tr}(W^\top P W). By Ky Fan's maximum principle, this trace is at most the sum of the kk largest eigenvalues of PP, with equality when WW spans the corresponding eigenvectors. Hence

1m∑dsd(W)≤1m∑i=1kλi(P).\frac{1}{m}\sum_{d} s_d(W) \le \frac{1}{m}\sum_{i=1}^{k}\lambda_i(P).

For mutually orthogonal subspaces, PP is a projector and the bound is k/mk/m, as in Lemma 1. The bound rises as the subspaces overlap, and it reaches kk only when all UdU_d coincide. A fixed filter therefore keeps a ceiling below kk whenever domains differ, and the orthogonal case is its least favourable instance.

The optimal subspaces on real EEG cannot be measured directly (see our reply to qvkh W3/Q2). Within the model of Lemma 1, however, the dimensions we use rule out the orthogonal case. In Rn\mathbb{R}^n, at most ⌊n/k⌋\lfloor n/k \rfloor subspaces of dimension kk can be mutually orthogonal, and any two of them share at least 2k−n2k - n dimensions (Grassmann formula). Our filters keep about half of the channels or more, so at most two domains can be mutually orthogonal. On BNCI2014001 and BNCI2015001 even two cannot be. On BNCI2014001 (k=20k = 20, n=22n = 22), any two domain subspaces share at least 18 of their 20 dimensions. Real domains therefore lie far from the worst case, and routing still improves over a fixed filter there (Table R2). This is consistent with the compression we observe. On the three datasets where we varied K, four experts already come within 1 point of K = m (Table R4). The revised PDF adds this bound next to Lemma 1 (Remark 1), and the dimension argument to Appendix F.

Q2 (is SCAP learning a per-domain bank; comparison with per-domain BiMaps). SCAP learns fewer shared experts than domains (Table R3, row domains / experts), and each domain has a preferred blend of them. The domain-routing alignment r_d confirms this. It compares how much the routing varies between domains with how much it varies across trials, and values near 1 mean the domain explains almost all of it. For input-routed SCAP it ranges from .77 to 1.11 (Table R1, row r_d), as in submitted Table 7, so routing is mostly driven by the domain. On BNCI2015001 and BNCI2014001, r_d is lower (.77 and .91), so routing also varies with the trial within a domain.

We added per-domain SPDNets, each trained on a single domain, as row (e) of Table R3. When the domain id is known, such per-domain baselines are the upper reference, and they use more parameters, 1.5 to 3.4 times SCAP's. Input-routed SCAP (c) is not significantly below them on Weibo2014 and BNCI2014001. They are significantly ahead on BNCI2015001 (+2.5 points), GrosseWentrup2009 (+5.2) and Schirrmeister2017 (+3.2). Against row (d), which learns one filter per domain with a shared classifier, SCAP is not significantly below on three of five datasets.

Q3 (sensitivity to K). Table R4. From K = 4 up to K = m, accuracy varies by at most 1.2 points on each dataset, and K = 2 already exceeds SPDNet on all three (+1.6 to +2.5 points). Tables R1–R3 use the paper's K, fixed before this sweep.

Reviewer_qvkhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the limited expressivity of stacked BiMap-ReEig layers in SPD networks. It combines an existing depth-collapse result with measurements showing that ReEig remains inactive on several preconditioned EEG datasets, and proves an overlap bound for a shared filter serving mutually orthogonal domain subspaces. SCAP replaces that filter with a mixture of Stiefel experts: a router uses covariance features and a domain embedding to combine tangent projections at a learned anchor, followed by a QR retraction. The paper gives a conditional local filter-bank approximation result, proposes mechanisms to reduce routing collapse, and evaluates synthetic examples and five binary motor-imagery EEG datasets. All evaluated domains contribute labeled training trials, and their identities are known at test time.

Critical strengths
  1. The connection between spectral closure and inactive ReEig layers gives a useful explanation for why adding BiMap-ReEig depth can fail.

  2. SCAP preserves SPD outputs while allowing the bilinear filter to depend on the sample and domain. The five-dataset comparison is encouraging: balanced accuracy improves over the plain BiMap baseline by 3.8 to 5.3 percentage points, with the reported comparisons surviving Holm correction.

Critical weaknesses
  1. Every test domain contributes labeled training trials, and domain identity is available at inference time. In this setting, a natural baseline is a bank of one learned BiMap filter per domain followed by a shared classification head. This is also the filter bank that Proposition 1 uses as its reference, yet the experiments omit it.

  2. The gains over SPDNet are not clearly attributable to dynamic routing. The results from table 6 and table 8 suggest that a substantial part of the improvement can arise from optimizing a larger parameterization of a fixed filter. SCAP uses 7.5 to 28.6 times as many trainable parameters as the baseline, making trained static controls and parameter-matched comparisons particularly important.

  3. Proposition 1 gives a local existence result under a realizable enclosing-simplex condition. The experiments do not estimate the affine dimension of the domain-optimal tangent filters, measure their distance from the anchor, or test the resulting approximation error.

  4. Appendix D describes per-domain whitening, variance equalization, recentering, and dimensionality reduction without specifying whether every fitted transform uses only the training partition of each fold.

  5. The five datasets provide variation in subjects, sessions, and recording conditions, but all experiments use binary motor-imagery classification. The results support effectiveness across several motor-imagery datasets without establishing whether the method generalizes to other EEG paradigms. This limits the empirical support for positioning SCAP as a general improvement to the expressivity of BiMap layers.

  6. Reducing the original multiclass tasks to binary classification also limits the evaluation of the proposed capacity improvement. The paper does not report performance with the full set of classes. For a method motivated by overcoming a capacity ceiling, multiclass classification would provide a useful test of whether the learned filters can preserve the discriminative information needed to distinguish several classes, beyond the distinctions required for a selected pair.

Questions for authors
  1. How much improvement remains after matching parameter count and optimization resources?

  2. What evidence supports filter-bank compression in the tested models?

3 comments
AuthorsOpenReview ↗

Response with pointers to revised article pdf

W1 (per-domain filter bank). We added this bank as row (d) of Table R3, and it is now in Table 1 of the revised PDF. Input-routed SCAP (c) is not significantly below it on Weibo2014, BNCI2015001 and BNCI2014001. The bank is significantly ahead on GrosseWentrup2009 (+6.8 points) and Schirrmeister2017 (+2.8). SCAP has slightly fewer parameters (row params). The bank, however, needs one filter per domain, while SCAP's number of experts is a free choice. Table R4 shows that K = 4 experts already come within 1 point of K = m, and SCAP's router needs no domain id.

W2 / Q1 (attribution to routing; parameter-matched comparison). This can be assessed with the uniform mixture of Table R2. It keeps the experts of input-routed SCAP and everything else in its training, and drops the router, which leaves it 77 to 99 % of the parameters. Input-routed SCAP exceeds it by +2.8 / +8.3 / +3.8 / +4.7 / +10.7 points (Weibo / GW09 / BNCI15 / BNCI14 / Schirr.; Holm p ≤ .015 on all five). The uniform mixture itself, with 4.6 to 8.5 times SPDNet's parameters, ranges from 3.3 points below SPDNet to 1.0 above. With nearly matched parameters, the gain over SPDNet therefore comes from routing.

W3 / Q2 (evidence for filter-bank compression). Compression is observed in the tested models. To show it, we varied the number of experts K (Table R4). With only K = 4 experts, input-routed SCAP is within 1 point of K = m, and not significantly below it, on all three datasets, which have 9, 18 and 28 domains. Proposition 1 itself is an existence result. It guarantees that K = r + 1 experts can match a per-domain bank to first order when the domain-optimal filters vary along r directions near the anchor. Measuring r, or the distance to the anchor, would require these optimal filters, which real data do not reveal, since filters trained on each domain only approximate them. The revised PDF therefore states Proposition 1 as a local existence result (Appendix F), and presents compression as an empirical observation (Section 7.3), without claiming that the near-anchor regime holds on real EEG.

W4 (how the preprocessing of App. D is fitted). We checked our code. Every fitted transform, namely the per-domain regularisation, whitening with variance equalisation, recentring, and dimensionality reduction where applicable, uses only the training partition, and is then applied unchanged to validation and test trials. The router's descriptor (standardisation and PCA) is also fitted on training trials only, as Appendix D now states. Our RPA implementation recentres each domain on all its trials without labels, which assumes the unlabelled test trials are available before classification. Appendix D now states this as well.

W5 / W6 (binary motor imagery only; full multiclass). We agree. All our experiments use binary motor imagery, and the revised PDF restates the scope accordingly (abstract, Section 7, Conclusion). It presents SCAP as an improvement for cross-domain binary motor-imagery decoding, and lists multi-class decoding and other paradigms (P300, SSVEP) as future work.

Thank you very much for your response and additional experiments. I think these results and clarifications have addressed my concerns about parameter-matched comparisons (W2), Proposition1 (W3) and preprocessing (W4).

I still have a question about W1. I acknowledge the revised router reduces reliance on an explicit domain id. However, Table R3 shows the per-domain filter baseline with a shared head achieves higher mean accuracy than SCAP on GW09, BNCI15, BNCI14, Schirr. Can you further explain under what conditions SCAP would be preferable to this simple baseline? What practical benefits does SCAP offer by reducing domain id dependency?

Thank you for the follow-up. We are glad that the parameter-matched comparison, Proposition 1 and the preprocessing are now clear.

EEG domains share much of their structure. Domains differ, so one fixed filter cannot serve them all, as routing adds 2.8 to 10.7 points over the uniform mixture (Table R2). Yet they also overlap strongly. In the model of Lemma 1, at the dimensions we use, any two domain subspaces share at least 18 of their 20 dimensions on BNCI2014001, and 11 of 12 on BNCI2015001 (Remark 1, Appendix F). Consistently, four experts come within one point of as many experts as domains, on datasets with 9, 18 and 28 domains (Table 3).

Why this matters for large datasets. Learning from large EEG datasets means training one model on many subjects and sessions, each of them a domain. A per-domain bank adds one n×kn \times k filter for every domain, so its size grows linearly with mm (1,800 parameters per domain on Weibo2014). SCAP's size is set by its KK shared experts, and in our sweep the number of experts needed did not grow with mm.

SCAP matches the bank with fewer parameters. SCAP rows come from Table 3, with four experts, or the paper's eight on BNCI2014001. Paired t over folds, Holm over the three datasets.

Weibo2014 (m = 9, K = 4)BNCI2014001 (m = 18, K = 8)BNCI2015001 (m = 28, K = 4)
bAcc (%), one filter per domain82.886.880.9
bAcc (%), SCAP82.385.280.5
difference (Holm p)−0.5 (.68)−1.7 (.32)−0.4 (.32)
parameters, SCAP / bank51 %55 %31 %

SCAP is statistically equivalent to the bank on all three datasets, with no significant difference and about half or less of its parameters.

Where SCAP is preferable. When one model must serve many domains that share structure, SCAP is equivalent to a per-domain bank, at a size that stays fixed as domains are added. This comes from the shared experts. Routing on the domain id would add only 8 parameters per domain, against 1,800 for a bank filter on Weibo2014. Routing on the trial removes even these at no significant cost (Table R3, rows b and c), as the router learns each domain's blend of experts from the signal itself (Appendix H). With few domains that lack a common basis, as on GrosseWentrup2009 and Schirrmeister2017 (Appendix D), a dedicated filter per domain remains the better choice.

Recent EEG models are trained on large corpora gathering many subjects and sessions, such as LaBraM (Jiang et al., 2024) and CBraMod (Wang et al., 2025). This trend motivates geometric models whose size stays fixed as the number of distributions they serve grows. SCAP is a step in this direction. Its experts lie on the Stiefel manifold, so every filter it builds preserves the SPD structure of the data, and the geometry remains built into the model as an inductive bias. Testing it on larger cohorts is future work.

Reviewer_mjMfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

SPDNet is a classical framework for learning from symmetric positive-definite (SPD) covariance matrices, including those used in EEG decoding. This paper addresses its limited gains from depth: when eigenvalue rectification (ReEig) is inactive, stacked bilinear mapping (BiMap) layers reduce to a single fixed spatial filter. This motivates more flexible filtering across subjects and sessions. The authors propose Stiefel Cross-Attention Pool (SCAP), which uses input- and domain-conditioned cross-attention to combine shared spatial-filter experts. Experts are blended in a tangent space around a learned anchor and mapped back onto the Stiefel manifold, yielding a sample-specific orthonormal filter while preserving SPD outputs. Three additional mechanisms mitigate routing collapse. The theory establishes a fixed-filter subspace-overlap ceiling for mutually orthogonal domain subspaces. It also shows that, when domain-optimal filters vary along few directions near a shared anchor, fewer experts than domains can approximate a dedicated per-domain filter bank to first order. On five EEG motor-imagery datasets with known subject/session domains, SCAP improves balanced accuracy over standard SPDNet by 3.8–5.3 percentage points, with less consistent advantages over domain-adaptive baselines. Overall, the paper proposes geometrically constrained adaptive filtering to expand SPDNet’s expressivity beyond stacking fixed BiMap layers.

Critical strengths

SCAP improves balanced accuracy over standard SPDNet across all five evaluated EEG datasets, with gains of 3.8–5.3 percentage points. The fixed-filter subspace-overlap bound and the local approximation result for a shared expert pool provide theoretical motivation for the design under stated assumptions. Together, these results support SCAP as a potentially useful extension of the SPDNet backbone.

Critical weaknesses

My main concern is that the current evidence does not sufficiently isolate which aspects of the proposed design are responsible for the improvements.

  1. Adaptive filtering versus improved optimization. In Table 6, naive SCAP performs almost identically when its learned experts are combined uniformly, suggesting that a multi-expert parameterization can improve performance even without adaptive routing. The full model does benefit from routing, but replacing its learned routing with uniform weights after training does not establish whether a fixed mixture could perform similarly if trained for that purpose. A comparison with the same expert pool trained from scratch using fixed, uniform mixing would help distinguish improved optimization from the benefit of adaptive filtering. This distinction is central to the claim of escaping the fixed-filter capacity ceiling.
  2. The added value of sample-dependent routing remains unclear. SCAP conditions its filters on both the current sample and domain identity. However, the query-masking results in Table 8 generally show greater sensitivity to removing domain identity than to removing sample features. Since the evaluated domains are represented during training, a domain-only router trained from scratch is an important comparison. This would clarify whether the improvements require trial-specific filtering or can largely be explained by learning a different fixed filter for each known domain.
  3. The benefit of mixing filters rather than their outputs is insufficiently established. SCAP combines experts on the Stiefel manifold before applying a single bilinear transformation. An alternative is to apply each expert to the input covariance matrix and then combine the resulting covariance matrices using the routing weights. Such a nonnegative weighted average also preserves positive definiteness. Comparing these approaches would clarify the value of the proposed filter-space construction beyond conditional expert weighting itself. Both predictive performance and computational cost would be relevant, since combining filters first could offer an efficiency advantage.
Questions for authors
  1. To what extent do SCAP’s gains come from adaptive routing rather than improved optimization through multi-expert parameterization? Naive SCAP’s comparable performance under uniform mixing makes this distinction important. A fixed-mixture counterpart trained independently would help establish whether adaptive filtering provides an additional benefit beyond learning a better fixed filter.
  2. How much does sample-dependent routing contribute beyond using domain identity? Table 8 suggests that removing domain information often has a larger effect than removing sample features, but test-time masking does not resolve this question. Comparison with a separately trained domain-only router would clarify whether trial-specific adaptation is important to the reported gains.
  3. What advantage is specific to combining filters on the Stiefel manifold? A natural alternative is to apply each expert separately and combine the resulting covariance matrices through routing weights, which also preserves positive definiteness. Comparing predictive performance and computational cost would clarify the value of SCAP’s filter-space construction beyond conditional expert weighting itself.
2 comments
AuthorsOpenReview ↗

Response with pointers to revised article pdf

W1 / Q1 (adaptive routing or better optimisation). We agree that swapping the learned weights for uniform ones at test time cannot separate routing from better optimisation. We therefore trained the control the reviewer describes, the uniform mixture of Table R2.

Routing accounts for the gain. The input-routed SCAP exceeds the uniform mixture by +2.8 / +8.3 / +3.8 / +4.7 / +10.7 points (Weibo / GW09 / BNCI15 / BNCI14 / Schirr.; Holm p ≤ .015 on all five). The two models share their experts and differ only by the router.

As the reviewer notes, the naive model of submitted Table 6 behaves as its own uniform-weight filter.

Correction. While preparing this control, we found an error in submitted Table 6. In the three rows with a learned anchor (Fix 1, Fix 1 + 2, Full), Δ_{K=1} mixed the experts around the first expert instead of W_base. Accuracies are unaffected. Recomputed on our rerun of the submitted SCAP (5 folds × 3 seeds), Δ_{K=1} is +0.4 / +4.6 / +5.5 / +4.6 / +10.0 points (Weibo / GW09 / BNCI15 / BNCI14 / Schirr.). In the revised PDF, the trained uniform mixture (Table R2) replaces this test-time proxy (Section 7.2), and the screen of the general response replaces submitted Table 6 (Appendix G).

W2 / Q2 (sample routing beyond the domain id). Table R3, rows (b) and (c). Trained from scratch, the domain-id router scores higher on three datasets, by at most 1.4 points, and the difference is not significant on any dataset. We keep the raw trial as the router input, since it is not significantly below the domain id on any dataset, without the id and without any per-domain parameter. As submitted Table 7 showed, each domain has a preferred blend of experts. SCAP builds this blend from fewer experts than domains (row domains / experts), and finds it from the trial alone.

W3 / Q3 (mixing filters vs mixing outputs). We trained the alternative the reviewer describes. Each expert is applied to the trial, and the K output covariances are averaged with the routing weights. Router, experts, initialisation and number of parameters are those of input-routed SCAP, which mixes the filters.

WeiboGW09BNCI15BNCI14Schirr.
SPDNet79.668.476.180.086.2
output mixing80.072.578.482.691.4
filter mixing (input-routed SCAP, proposed)83.374.780.285.293.5
filter − output mixing (Holm p)+3.4 (.052)+2.2 (.137)+1.9 (.085)+2.6 (.052)+2.2 (.137)
training time per epoch (s), output mixing0.642.460.860.556.04
training time per epoch (s), filter mixing0.692.590.870.596.47

Filter mixing scores 1.9 to 3.4 points higher on all five datasets. The difference has the same sign everywhere, but does not reach significance after Holm correction. Output mixing also improves over SPDNet, so routed experts help in both forms, and mixing them in filter space adds to this gain. Training costs about the same, with output mixing 1 to 7 % faster per epoch in our implementation. At inference, filter mixing applies a single bilinear map per trial, plus the QR of the mixed filter, instead of K bilinear maps.

Thank you to the authors for conducting the requested control experiments across multiple seeds.

Several of the new results help clarify the behavior of the framework. In particular, Table R2 confirms that learned routing significantly outperforms a uniform mixture trained from scratch (addressing W1/Q1), and the comparison in W3/Q3 shows a consistent positive margin (+1.9 to +3.4 points across all five datasets) for filter-space mixing over output-covariance mixing, even though the individual differences fall just short of significance after Holm correction (Holm p in [0.052, 0.137]).

However, two central issues remain:

  1. Substantial change to the core method during review: Based on the multi-seed screen and the correction to Table 6, the authors propose removing all three structural mechanisms that formed the core technical narrative of Sections 5, 6, and Appendix G—namely, the learned tangent-space anchor W_base (Fix 1), the domain-separating projection and domain embedding (Fix 2), and the winner-take-all key alignment loss (Fix 3)—in favor of ambient-space averaging with QR retraction and cross-entropy training alone. While this simplification is well-motivated by the new screen, it invalidates much of the submitted paper's methodological exposition, collapse diagnostics, and ablation analysis, requiring a major rewrite beyond the scope of a standard revision.
  2. Limited evidence for trial-specific adaptation beyond domain conditioning (W2/Q2): Table R3 shows that a router conditioned solely on the domain ID (row b) matches or slightly exceeds the proposed trial-routed model (row c) on four of the five datasets. Moreover, because all evaluated methods still require known domain groupings at test time for per-domain recentering, feeding the un-recentered trial covariance to the router primarily allows it to recover between-domain shifts (r_d in [0.77, 1.11]) rather than demonstrating a clear benefit from trial-specific filtering within a domain.

Because the rebuttal results entail replacing the main technical mechanisms of Sections 5-6 and show no accuracy advantage of trial-level routing over domain-level routing, I do not think the manuscript is ready for acceptance in its current form and will maintain my score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response: a simpler SCAP

We thank the three reviewers for their careful reading. We are glad that they found value in the diagnosis and in the theory. Inactive ReEig layers explain why BiMap–ReEig depth fails to help (qvkh, srQU), and the fixed-filter ceiling of Lemma 1 is easy to understand and motivates the design (srQU, mjMf). All three reviewers also noted that SCAP improves over SPDNet on all five datasets.

We uploaded a revised PDF, with changes in blue. SCAP is one model shared by all domains, which adapts its filter to each domain with fewer experts than domains. As submitted Table 7 already showed, its routing is mostly explained by the domain. Each domain has a preferred blend of experts, and individual trials vary around it. The reviewers' questions led us to examine which components produce the gains. From this analysis, we concluded that a simpler layer suffices. Its router reads only the trial, without the domain id, and its blends still follow the domain, with some variation across trials. This input-routed SCAP, which the revised PDF now presents (Sections 5–6), matches or exceeds the submitted one on all five datasets, with fewer parameters, and four experts come within one point of using as many experts as domains, on datasets with up to 28 domains (reply to qvkh W3/Q2). When the domain id is known, the upper reference is a per-domain filter bank, which learns one filter for each domain and therefore more filters than SCAP has experts. Input-routed SCAP is not significantly below it on three of five datasets (reply to qvkh W1).

To examine the components, we screened 120 variants of the layer on Weibo2014, BNCI2014001 and BNCI2015001 (5 folds × 2 seeds each), and selected one on validation scores only. We first compared router inputs. Averaged over the screen, the trial's covariance before recentring is as good a routing signal as the domain id (82.6 % test bAcc for both). It carries the routing information that the id provides, and it also lets trials of the same domain be routed differently. We therefore route on the raw trial alone. This replaces both parts of the submitted router input (Fix 2), the projection of the recentred trial and the domain embedding, so the router no longer uses the domain id. The two other fixes, the learned anchor (Fix 1) and the alignment loss (Fix 3), change accuracy by less than half a point on average over the screen, for every router input, the submitted one included. Mixing around a learned anchor scores 0.1 points below ambient mixing followed by a QR retraction, and the alignment loss 0.3 points below cross-entropy alone. Submitted Table 6 measured these fixes on Weibo2014 with a single seed, where their effects reached 1.3 points. Averaged over three datasets and two seeds, they are smaller, so we keep the simpler options. To first order, ambient mixing equals tangent mixing around an implicit anchor, so Proposition 1 still applies (derivation in Appendix B). In summary, input-routed SCAP routes on the trial's covariance before recentring, without the projection or the domain embedding (Fix 2). It mixes the experts in the ambient space and retracts with a QR decomposition, without a learned anchor (Fix 1), and it is trained with cross-entropy alone, without the alignment loss (Fix 3).

All arms below were retrained on identical folds and averaged over 3 seeds, where the paper reported a single seed, so values differ from submitted Tables 1–2. Every arm uses the same preprocessing, and every neural arm the same optimiser and selection rule. Each domain is recentred on its training trials, SPDNet included (reply to qvkh W4).

Paired t-test over folds (seeds averaged per fold), Holm correction over the five datasets.

Table R1: submitted vs input-routed SCAP.

WeiboGW09BNCI15BNCI14Schirr.
bAcc (%), SPDNet79.668.476.180.086.2
bAcc (%), submitted SCAP81.172.679.883.790.4
bAcc (%), input-routed SCAP (proposed)83.374.780.285.293.5
params (k), submitted / input-routed20.7 / 15.956.7 / 48.69.0 / 3.516.7 / 4.672.0 / 63.9
H̄, submitted / input-routed.725 / .694.670 / .476.513 / .667.577 / .739.584 / .249
r_d, submitted / input-routed.688 / 1.107.971 / 1.099.822 / .770.751 / .911.999 / 1.030

Input-routed SCAP scores 0.4 to 3.2 points above the submitted SCAP, significantly on BNCI2014001 (Holm p = .027). Its routing differs between domains (r_d from .77 to 1.11; as in submitted Table 7, r_d can slightly exceed 1), and within a trial the weight concentrates on a few experts (H̄ from .25 to .74).

AuthorsOpenReview ↗

General response: a simpler SCAP .2

Tables R2–R4 complete the general response. They follow the same protocol and tests as Table R1, and the replies to each reviewer refer to them.

Table R2: uniform mixture trained from scratch. The uniform mixture is the input-routed SCAP with α = 1/K fixed during training and test, and no router. It shares everything else with the routed model, from the experts and their initialisation to the training schedule, folds and seeds.

WeiboGW09BNCI15BNCI14Schirr.
bAcc (%), SPDNet79.668.476.180.086.2
bAcc (%), uniform mixture80.666.476.480.582.9
bAcc (%), input-routed SCAP (proposed)83.374.780.285.293.5
params (k), SPDNet / uniform / input-routed2.7 / 15.3 / 15.97.6 / 48.0 / 48.60.3 / 2.7 / 3.50.9 / 3.9 / 4.67.6 / 63.2 / 63.9

Table R3: per-domain comparisons. The domain-id router (b) is the input-routed SCAP whose router reads only a learned embedding of the domain id, trained from scratch. One filter per domain, shared head (d) learns one BiMap per domain, followed by a shared classifier. One SPDNet per domain (e) trains a full SPDNet on each domain alone.

WeiboGW09BNCI15BNCI14Schirr.
domains / experts9 / 810 / 928 / 1618 / 813 / 12
(a) SPDNet79.668.476.180.086.2
(b) domain-id router83.076.181.585.593.5
(c) input-routed SCAP (proposed)83.374.780.285.293.5
(d) one filter per domain, shared head82.881.580.986.896.3
(e) one SPDNet per domain82.679.982.886.496.8
params (k), (c) / (d) / (e)15.9 / 17.1 / 24.648.6 / 53.1 / 76.03.5 / 4.5 / 8.84.6 / 8.3 / 15.563.9 / 68.2 / 98.8

Rows (a)–(e) are bAcc (%).

Table R4: number of experts. Input-routed SCAP trained with K experts on Weibo2014, BNCI2014001 and BNCI2015001 (5 folds × 3 seeds). Tables R1–R3 use the paper's K, fixed before this sweep.

bAcc (%)SPDNetK = 2K = 4K = 8K = m
Weibo (m = 9)79.681.282.383.383.0
BNCI14 (m = 18)80.082.584.085.284.9
BNCI15 (m = 28)76.178.580.580.680.9

Changes in the revised PDF (in blue). Section, table and appendix numbers in our replies refer to the revised PDF; "submitted Table 6" and similar refer to the original. The method change is confined to the three choices summarised in the first part of this response (Sections 5.1 and 6). The theory (Lemma 1, Proposition 1) and the architecture of the layer are unchanged, and the other changes answer the reviewers' requests.

  • SCAP routes on the trial before recentring, without the domain id, mixes its experts in the ambient space and is trained with cross-entropy alone (Sections 5–6, Fig. 1). Appendix B gives the first-order derivation, and Appendix G the selection screen, which replaces submitted Table 6.
  • Remark 1 extends the ceiling to overlapping subspaces. Proposition 1 is stated as a local existence result, with the dimension argument (Appendix F).
  • Table 1 adds MUSUME and one filter per domain, with 3 seeds and Holm-corrected tests. Table 2 adds the controls (uniform mixture, output mixing, routing by domain id), and Table 3 the number of experts. Appendix H gives the routing diagnostics and the per-domain references.
  • Appendix D states that every fitted transform uses training trials only, and that our RPA recentres each domain on all its trials.
  • The scope is restricted to binary motor imagery, and the conjecture on medical imaging and radar is removed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.