acceptodds
Under review as a conference paper at ICLR 2027

CogEEG: Cognition-Guided Hyperbolic Learning for Cross-Subject EEG–Visual Retrieval

Abstract

Cross-subject EEG–visual retrieval requires learning from source subjects while generalizing to unseen subjects. Existing methods typically perform subject alignment on a global uniform representation. However, such a level-agnostic paradigm overlooks the heterogeneity across cognitive levels (i.e., different levels exhibit distinct EEG–visual correspondence patterns and varying degrees of cross-subject distribution shift), thereby limiting cross-subject generalization. To address this limitation, we propose CogEEG, a Cognition-guided hyperbolic learning framework for cross-subject EEG–visual retrieval from global uniform alignment to cognitive-level-aware adaptive alignment. Specifically, CogEEG builds multi-level EEG representations from brain-region priors and captures their hierarchy via level-specific hyperbolic geometry and geodesic cross-level interaction. These representations are further aligned with multi-level visual features to establish structured EEG–visual correspondences. Building on the hierarchical representations, CogEEG dynamically adjusts domain-adversarial alignment across cognitive levels based on visual reliability and subject discriminability, enabling level-aware cross-subject alignment. This alignment reduces level-specific subject shifts while improving generalization to unseen subjects. Extensive experiments show that CogEEG achieves superior performance on cross-subject EEG–visual retrieval.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
EEG–visual retrieval, EEG–visual correspondences, Cross-subject generalization

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

The authors target the problem of visual retrieval using EEG data, specifically in cross-subject evaluation. They argue that current approaches don't fully exploit a hierarchical relationship between the cognitive levels involved in image retrieval. As such, they recommend the use of hyperbolic geometry that has been shown to better represent hierarchies. To do this the authors extend ATM-S with two modules. Brain-Region Guided Hierarchical Cognitive Pathway (BHCP) splits electrodes into 3 separate groups and treats them as low, middle and high cognitive levels. Each level is projected onto the hyperbolic manifold with a learnable curvature and then merged into one, and the levels interact through attention where the weights are a softmax over the Poincare distances. Two experts are then formed as the weighted Frechet means of the low-middle and middle-high geodesics. The experts are supervised by shallow and deep CLIP features, and they're fused by a geodesic interpolation based on their hyperbolic distance and radii. The fused point is mapped back to Euclidean space and added to the ATM-S embedding, so retrieval itself is cosine similarity in Euclidean space. The second component, Adaptive Hierarchical Domain-Adversarial Learning (AHDA) applies a per-level domain-adversarial subject classifier to the Euclidean tangent features, gives weights to the three levels based on a softmax over moving averages of visual reliability and subject discriminability, and then projects adversarial gradients that conflict with retrieval. They then test the method on two datasets in a 200-way zero shot setting.

Critical strengths
  • LOSO settings in general in EEG problems are relatively weak and targeting helps progress the field.
  • The ablation studies in principle are comprehensive but need multi-seed verification.
Critical weaknesses
  • One of my major concerns with the paper is the presentation of the results. To begin with, the experiments are single seed runs without standard deviations or significance tests. If we have 200 test queries per subject, the component effects in the ablations are made less significant since a 0.5 difference is roughly only a single image. This makes multiple seeds more important to better estimate actual performance.
  • Adding on that, [1] is cited in the paper as hyperbolic work on EEG but it is never compared against despite it also doing EEG-to-CLIP alignment on the same THINGS-EEG 200-way protocol with learnable curvature, a perceptual/semantic split and learned geodesic interpolation. They also seem to report higher scores overall.
  • The authors claim that the EEG sensor groups represent different cognitive levels but the only support given is a single citation (Karimi-Rouzbahani 2018), which works with peri-occipital and peri-frontal electrode groups over time windows rather than a three-level regional hierarchy. This is supposed to motivate the use of hyperbolic geometry so it would be important to justify properly. The standard view (including Felleman and Van Essen, Grill-Spector and Malach, and DiCarlo et al., all cited) puts high-level object representation in ventral temporal cortex, not fronto-central sites. Frontal electrodes also carry the most EOG, and the preprocessing has no ocular correction. The paper's own data agrees when they find that the frontal alone gives 2.0 Top-1 and central 3.1 (Table 3), and the high Shapley row is 0.007-0.023 against 0.283 for low (Fig. 4). The channel lists for each EEG group are also never given, and no ablation drops the fronto-central branch.
  • Some components aren't attributed properly like Eq. 7, which is hyperbolic distance attention [2, 3, 4], and [4] already has Poincare multi-head attention. It's also not described how the value aggregation is done, and the hyperbolic linear Q/K/V maps are never defined. The gradient projection in Eq. 16 is the PCGrad projection [5] and is not cited. Additionally some of the AHDA components are in the literature as per-level adversarial losses [6], weaker global alignment [7], and entropy-based transferability weighting [8], which has the same 1 - entropy form as U_l.
  • Missing hyperbolic and manifold EEG literature such as HEEGNet [9], LAtte [10], MAtt [11] and GyroAtt [12]. *Lastly, the paper suffers from heavy AI formatting including excessive italics and dashes and it leads to some formatting issues as well like "Poincar'e" everywhere. There's other issues like Fig. 3 uses five regions while the method uses three, and "NICE-G" and "Cog-cap" are never defined etc... These need cleaning and corrections

[1] S. Jo, W. Jeong, D.-W. Heo, Y. Hwang, H.-I. Suk. "HyFI: Hyperbolic Feature Interpolation for Brain-Vision Alignment." [2] C. Gulcehre, M. Denil, M. Malinowski, A. Razavi, R. Pascanu, K. M. Hermann, et al. "Hyperbolic Attention Networks." [3] W. Chen, X. Han, Y. Lin, H. Zhao, Z. Liu, P. Li, M. Sun, J. Zhou. "Fully Hyperbolic Neural Networks." [4] R. Shimizu, Y. Mukuta, T. Harada. "Hyperbolic Neural Networks++." [5] T. Yu, S. Kumar, A. Gupta, S. Levine, K. Hausman, C. Finn. "Gradient Surgery for Multi-Task Learning." [6] Y.-H. Tsai, W.-C. Hung, S. Schulter, K. Sohn, M.-H. Yang, M. Chandraker. "Learning to Adapt Structured Output Space for Semantic Segmentation." [7] K. Saito, Y. Ushiku, T. Harada, K. Saenko. "Strong-Weak Distribution Alignment for Adaptive Object Detection." [8] X. Wang, L. Li, W. Ye, M. Long, J. Wang. "Transferable Attention for Domain Adaptation." [9] S. Li, S. Chu, O. Koc, Y. Ding, Q. Zhao, M. Kawanabe, Z. Chen. "HEEGNet: Hyperbolic Embeddings for EEG." [10] J. Burchert, A. Bdeir, T. Hanika, L. Schmidt-Thieme, N. Landwehr. "LAtte: Hyperbolic Lorentz Attention for Cross-Subject EEG Classification." [11] Y.-T. Pan, J.-L. Chou, C.-S. Wei. "MAtt: A Manifold Attention Network for EEG Decoding." [12] R. Wang, C. Hu, X. Song, X.-J. Wu, N. Sebe, Z. Chen. "Towards a General Attention Framework on Gyrovector Spaces for Matrix Manifolds."

Questions for authors
  1. Are there 200 or 4,000 test queries per subject? The reported values are multiples of 0.5 which correspond to 200 so that might be a typo.
  2. What is the "hyperbolic linear transformation" for Q/K/V, and how are values aggregated?
  3. Why are the expert weights fixed at 1/3 and 2/3? How sensitive is the model to these values?
Reviewer_hZnGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CogEEG for cross-subject EEG–visual retrieval. The method contains two main components. BHCP constructs hierarchical EEG representations from predefined brain-region groups, models cross-level interactions in hyperbolic space, and aligns perceptual and semantic experts with multi-level CLIP features. AHDA applies adaptive domain-adversarial alignment based on visual reliability and subject discriminability, together with retrieval-aware gradient projection. The authors evaluate CogEEG on THINGS-EEG and THINGS-MEG under intra-subject and inter-subject settings, reporting positive improvements over the selected baselines, especially ATM. Although the problem is meaningful and the reported improvements have empirical value, several important concerns remain. The cognitive-level assumptions are not sufficiently validated, the novelty over closely related prior works is unclear, and the experiments provide limited evidence for some key technical claims. These issues substantially weaken the scientific contribution and the conclusions of this work.

Critical strengths
  1. A structured design for hierarchical EEG–visual alignment. BHCP goes beyond simply separating EEG channels into regional branches. It uses three STGCN encoders (Sec. 3.2, Eq. 6), introduces perceptual and semantic experts with a shared middle-level representation (Eq. 8), and applies multi-level CLIP supervision to differentiate their roles (Eqs. 9–10). The geometry-aware expert fusion (Eqs. 11–12) also provides an explicit mechanism for combining these representations. These are concrete and reasonably motivated design choices for studying region-dependent EEG–visual alignment.

  2. Consistent empirical improvements across held-out subjects. Table 1 shows that CogEEG improves the average inter-subject Top-1 accuracy from 9.90% to 14.25% over the reproduced ATM baseline. Importantly, the improvement is observed across all ten held-out subjects, rather than being driven by a few subjects. The leave-one-subject-out protocol also provides a relevant evaluation setting for cross-subject generalization (Appendix B.2). Table 2 further reports positive Top-1 gains across all four subjects on THINGS-MEG, although the average improvement is relatively small.

  3. Useful analyses of regional information and hierarchical alignment. The paper provides several experiments that help examine its motivation. Table 3 shows that occipital channels achieve 31.80% Top-1, while all channels reach 38.50%, suggesting additional information from other regions. Fig. 4 uses cross-validated Shapley contributions to study EEG–visual correspondence across levels. Table 4 further examines the effects of aligning different levels, and Tables 5–8 analyze the proposed modules and simpler alternatives. These analyses are useful for understanding the method, even though some findings also raise concerns about the authors' interpretation.

Critical weaknesses
  1. The main cognitive-level assumptions are not sufficiently supported. The Introduction suggests that global alignment implicitly assumes similar subject shifts across cognitive levels, but this does not necessarily follow from using a global representation. More importantly, BHCP directly associates three spatial EEG channel groups with low-, middle-, and high-level cognitive processing (Sec. 3.2, Eq. 6). This is a reasonable engineering prior, but scalp EEG channels do not directly isolate distinct cognitive levels, considering neural source mixing and the temporal dynamics of visual processing. Although the encoder models temporal patterns, the proposed hierarchy is mainly defined by spatial groups. Fig. 4 also shows relatively weak high-level correspondence, while Table 4 does not directly measure level-specific subject shifts. The neuroscience motivation therefore needs more direct evidence.

  2. The novelty over closely related prior works is not clearly established. Several recent works have already studied closely related ideas. ViEEG (Liu et al., ICML 2026) explores hierarchical visual representations and multiple EEG streams. HyFI (Jo et al., AAAI 2026) uses hyperbolic geometry for perceptual–semantic brain–vision alignment. RAHS (Zhu et al., CVPR Findings 2026) investigates region-aware hierarchical EEG–visual alignment, while HEEGNet (Li et al., ICLR 2026) studies hyperbolic EEG representations and cross-domain generalization. SAMGA (Jiang et al., arXiv 2026) further explores subject-aware multi-granularity visual supervision. Although some of these works are cited, Sec. 2 does not sufficiently explain the substantive differences of CogEEG. Table 1 also mainly compares older baselines, without controlled comparisons with stronger recent approaches. This makes the additional methodological novelty and empirical competitiveness difficult to assess.

  3. The experimental evidence for several core technical contributions is relatively weak. Although the full framework improves over ATM, the additional gains of some proposed mechanisms are small. Adaptive alignment improves Top-1 by only 0.30 percentage points over uniform alignment (Table 7). Hyperbolic geometry improves Top-1 by 0.55 points over Euclidean geometry in the intra-subject setting (Table 8), while gradient projection adds only 0.25 points (Appendix Table 10). Their corresponding Top-5 results are also lower. Moreover, Table 5 shows that BHCP alone reaches 14.00% and AHDA alone reaches 13.95%, compared with 14.25% for the full model. This suggests that much of the improvement is already achieved by simpler configurations. Without multi-seed results or uncertainty estimates, the independent benefits and stability of these mechanisms remain insufficiently demonstrated.

  4. The role of level-specific hyperbolic geometry is unclear. In Sec. 3.2, regional EEG representations are first mapped into level-specific Poincaré spaces and then transferred through the tangent space into a shared Poincaré space. Based on the presented formulation (Eqs. 5–7), the level-specific exponential and logarithmic mappings appear to cancel before the shared-space interaction. The resulting representations therefore seem to depend on the average curvature rather than the individual level-specific curvatures. This raises a concrete concern about whether independently learned curvatures actually provide the claimed level-specific geometric modeling. The authors should clarify the computation and the actual role of these parameters.

  5. Some important experimental protocol and implementation details remain unclear. Appendix A.1 states that multivariate noise normalization is estimated from training data and applied separately for each participant. However, it is unclear how the normalization statistics are obtained for a completely unseen test subject. The main symmetric retrieval loss is also not described in enough detail to determine how multiple EEG trials corresponding to the same image are handled, although Eq. 9 explicitly considers unique images. In addition, the AHDA-only ablation in Table 5 needs a clearer explanation of which hierarchical representations remain available when BHCP is removed. The absence of released code further limits independent verification of these details.

Questions for authors
  1. Could the authors provide more direct evidence that the three spatial EEG groups represent distinct cognitive processing levels? For example, would alternative channel groupings or time-resolved EEG–visual analyses support the same hierarchy? Can the differences in subject shifts across these levels also be measured directly?

  2. What are the main technical and scientific differences between CogEEG and closely related works such as ViEEG, HyFI, RAHS, HEEGNet, and SAMGA? Could the authors explain the choice of baselines and whether stronger recent methods can be evaluated under a matched protocol?

  3. Are the relatively small improvements from adaptive alignment, hyperbolic geometry, and gradient projection consistent across multiple random seeds? Could the authors also explain the decreased Top-5 results in several ablations and the limited additional gain from combining BHCP and AHDA?

  4. Could the authors clarify how independently learned level-specific curvatures affect the final representations, given the apparent cancellation of the exponential and logarithmic mappings in Sec. 3.2? Also, how are normalization statistics obtained for the unseen test subject in the leave-one-subject-out protocol?

References

Liu et al. ViEEG: Hierarchical Visual Neural Representation for EEG Brain Decoding. ICML, 2026.

Jo et al. HyFI: Hyperbolic Feature Interpolation for Brain-Vision Alignment. AAAI, 2026.

Zhu et al. Region-Aware Hierarchical Sub-Feature Alignment for Robust EEG-Based Visual Decoding. CVPR Findings, 2026.

Li et al. HEEGNet: Hyperbolic Embeddings for EEG. ICLR, 2026.

Jiang et al. Subject-Aware Multi-Granularity Alignment for Zero-Shot EEG-to-Image Retrieval. arXiv:2604.17782, 2026.

Reviewer_QK7BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CogEEG combines regional hyperbolic EEG representations with adaptive subject-adversarial alignment for cross-subject image retrieval, evaluated on THINGS-EEG and THINGS-MEG.

Critical strengths

The improvement over reproduced ATM is substantial: inter-subject Top-1/Top-5 rises by 4.35/8.20 points, with Top-1 gains for all ten EEG subjects. The regional analyses and component controls are useful.

Critical weaknesses
  1. The claimed level-specific geometry is not supported by the written computation. In Section 3.2, Log⁡0cℓ(Exp⁡0cℓ(tℓ))=tℓ\operatorname{Log}^{c_\ell}_0(\operatorname{Exp}^{c_\ell}_0(t_\ell))=t_\ell, so the transfer reduces to zℓ=Exp⁡0cˉ(tℓ)z_\ell=\operatorname{Exp}^{\bar c}_0(t_\ell). Individual curvatures affect subsequent processing only through their mean. This can still be a useful shared-curvature model, but the proposed geometric explanation needs revision unless an operation is missing.

  2. Adaptive alignment is not clearly better than the existing Uniform control. Table 7 shows +0.30 points in Top-1 but −1.60 in Top-5; Table 5 likewise shows small, mixed component gains. Without training-repeat uncertainty, these results do not convincingly establish the claimed adaptivity or complementarity.

  3. The regional analyses do not separate the cognitive partition’s benefit from the additional branches and model capacity.

Questions for authors

Is there an operation between the level-specific Exp and Log maps that is omitted from the description?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.