- 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
- 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.
- What is the "hyperbolic linear transformation" for Q/K/V, and how are values aggregated?
- Why are the expert weights fixed at 1/3 and 2/3? How sensitive is the model to these values?