acceptodds
Under review as a conference paper at ICLR 2027

LEARNING SUBJECT STABLE CORTICAL DYNAMICS FROM EEG WITH EVENT SYNCHRONIZED INVARIANT MODES

Abstract

EEG based motor decoding is usually built around a separation that is convenient computationally but weak scientifically. Source imaging is used to estimate where cortical activity occurs, while classification is applied afterward to decide what movement was performed. For fine upper limb actions, this separation is problematic because the most anatomically complete reconstruction is not necessarily the representation that transfers best across people. We introduce Event Synchronized Invariant Cortical Modes, ESICM, which learns cortical patterns that are simultaneously tied to the forward model, aligned with movement timing, discriminative across actions, and stable across subjects. Instead of reconstructing a dense source map and then compressing it for prediction, ESICM learns a small set of cortical modes directly from training data through a leadfield constrained generalized eigen problem with subject variability and cortical smoothness penal- ties. Multiband EEG is projected into these modes and decoded by a lightweight temporal network, producing a compact cortical representation without a parallel sensor branch or post hoc sensor source fusion. All supervised operations are con- fined to the training subjects under nested leave one subject out evaluation, with classical inverse methods retained as anatomical references. The resulting framework treats source structure, cross subject stability, and efficient decoding as a single representation problem, providing a direct path from scalp EEG to compact cortical dynamics suitable for embedded brain computer interfaces.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG decoding, cortical representation learning, cross-subject generalization, brain-computer interface, invariant representation learning

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

Authors propose a method that learns a subset of spatial modes per frequency band by solving general eigen problem.

This combines leadfield, class contrast and subject instability operator and graph laplacian penalty. The projections feed a depthwise separable network with cosine prototype head.

Eval is done on 15 subjects with Jetson nano latency measurements.

Critical strengths

Data leakage is well controlled. benchmarks 20 sensors and source configurations and reports end to end embedded latency which is good for BCI deployment.

Critical weaknesses
  1. this is not jointly learned localization as the modes come from a closed-form eigensolution that is frozen before decoder training.

  2. controlled comparison does not isolate representation as ESICM gets 4 band views and qt in both mode estimation and decoder. while the baselines get neither and there is no Gw = I ablation.

  3. table 7 does not measure invariance as LOSO folds share 13 of 14 training subjects, and there is no CSP/MNE reference subspace and no null for motor concentration.

  4. the scope is 2 cue-locked datasets of 15 and 106 subjects with no asynchronous evaluation which fits a JBHI-EMBS or CCNeuro type venue better than ICLR.

Questions for authors
  1. what happens to accuracy with G_w = I, and with the same four views and event-conditioned decoder applied to sensor and MNE inputs?

  2. do paired Wilcoxon tests support the reported margins across all 8 settings and on BCI2000?

  3. q(t) is dominant in ablation. what will the results show with a fixed class-agnostic window instead of the population-derived q(t), and how would q(t) be obtained in asynchronous decoding?

I like the paper from a computational neuroscience and BCI framework, but not from an applied ML in neuroscience standpoint. There seems to be no ML novelty.

can the authors justify this submission to ICLR? I am open to changing my mind.

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

The paper focuses on cross-subject EEG-based motor decoding. It questions the suitability of reconstructed source representations and the separate optimization of source reconstruction and classification for this task. It proposes ESICM, which learns compact cortical modes in place of dense source reconstruction and trains a decoder on their time courses. Evaluation on two EEG datasets examines the decoding accuracy of the combined representation and decoder, component contributions through ablations, and deployment efficiency.

Critical strengths

Empirical support for cross-subject decoding. The accuracy comparisons on two EEG datasets under subject-disjoint protocols provide some empirical support for the complete ESICM system’s decoding performance, while the ablations in Section 3.2 suggest that multiple components contribute to this performance.

Critical weaknesses
  1. The core problem and its bottleneck are insufficiently established. Cross-subject EEG-based motor decoding provides the problem context, while the paper’s central criticism concerns the suitability of reconstructed source representations and the separate optimization of source reconstruction and classification. However, the Introduction does not sufficiently explain how existing sensor-space and source-space methods are limited in cross-subject decoding, why the focus shifts to source reconstruction, or what specific limitation separation introduces. It then motivates ESICM through fine-grained movement discrimination, while the Abstract concludes with a claim about unified representation learning. These shifts leave the connection between the core problem, the targeted bottleneck, and the claimed contribution unclear, making the scientific value of the proposed approach difficult to assess.
  2. The experimental evidence is insufficient for the central claims. The experiments do not adequately establish how separation limits cross-subject decoding or demonstrate the advantages of the representation itself. The main accuracy comparisons change both the representation and the decoder, and the ablations in Section 3.2 suggest that decoder design may also contribute to the reported gains through event-conditioned decoding. They therefore do not sufficiently isolate the representation’s contribution to the reported improvements. Furthermore, Appendix Table 7 compares population-level mode bases learned across LOSO folds, rather than representations of unseen subjects. It does not directly establish cross-subject representation invariance. The reported system-level performance thus does not sufficiently validate the claimed representation benefits, particularly invariance and transferability.
Questions for authors

What specific bottleneck in cross-subject decoding does ESICM aim to address?

How does it relate to reconstructed source representations and the separate optimization of reconstruction and classification?

Which aspect of it does ESICM address, and how do the experimental results support that claim?

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

This work learns a small set of spatial filters per frequency band by solving a CSP-style generalized eigenproblem in a cortical source space. The denominator adds a subject-instability term, a cortical Laplacian, and a ridge through a whitened leadfield, and trial covariances are weighted by an event field estimated on training subjects. The projections feed a small temporal decoder with prototype classification. Under nested LOSO on the Ofner upper-limb dataset, ESICM is reported to achieve the highest mean accuracy in all eight motor-execution and motor-imagery settings against LDA, MDM, CNN, EEGNet, and Transformer decoders on sensor and source-space inputs, with an ablation study, a four-class BCI2000 experiment, paired tests for one setting, and about 4 ms end-to-end latency on a Jetson Nano.

Critical strengths
  1. The protocol is careful. It uses nested LOSO with all statistics estimated on training subjects, paired Wilcoxon tests with Holm correction, cross-fold reproducibility of the modes, and a second dataset. Not comparing against published numbers from leaky protocols is the right decision.
  2. Combining discrimination, subject invariance, and spatial smoothness in one closed-form eigenproblem is a compact idea.
Critical weaknesses
  1. The method cannot be reproduced from the paper. No values or search ranges are given for the number of modes, α, β, ρ, ε, the head model (Ofner has no MRI), the source space, or the decoder; the text only says that hyperparameters are chosen on inner folds, and the rest is deferred to a code release. "CNN" and "Transformer" in the tables are not specified, and with a single figure the learned modes are never shown.
  2. Whether source space is needed is not tested. Table 3 removes the event weighting, the subject term, and the Laplacian, but never the leadfield. With a full-rank leadfield the projection spans all of sensor space, so without the Laplacian the method is a sensor-space regularized CSP with a subject term, close to [1]. No CSP or FBCSP baseline is included, and the usual cross-subject recipe of Euclidean alignment [2] plus EEGNet is not run.
  3. Some baselines look misconfigured. MDM is at chance on BCI2000 and EEG-DANN is far below the plain CNN on the same inputs (Table 5), so the argument that decoder capacity cannot close the gap rests on these numbers, and the margins over Sensor+CNN are a few points with a paired test for one of eight settings. Algorithm 1 solves the eigenproblem and then trains the decoder, so this is not the "single representation-learning problem" claimed, and sensor EEGNet is faster and smaller (Table 4).

[1] Lotte F, Guan C. Regularizing Common Spatial Patterns to Improve BCI Designs: Unified Theory and New Algorithms. IEEE Transactions on Biomedical Engineering, 58(2):355-362, 2011.

[2] He H, Wu D. Transfer Learning for Brain-Computer Interfaces: A Euclidean Space Data Alignment Approach. IEEE Transactions on Biomedical Engineering, 67(2):399-410, 2020.

Questions for authors
  1. What are the values of K, α, β, ρ, ε, the head model, and the atlas, and how were they selected?
  2. Does the same eigenproblem with an identity leadfield and a sensor-space smoothness prior reach the same accuracy?
  3. How does ESICM compare with Euclidean alignment [2] plus EEGNet and with a regularized CSP baseline [1] under the same protocol?
  4. Why are MDM and EEG-DANN so far below the plain CNN?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.