- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper proposes a divergence between two constant-curvature spaces with potentially different curvatures by extending Sliced Gromov Wasserstein to spherical spaces and combining it with its versions in Euclidean and hyperbolic spaces. The authors provide theoretical bounds relating the proposed divergence with intrinsic GW. The paper also provides experiment results to show this divergence is effective and could be beneficial as a regularizer when incorporated with learning methods on graph and multimodal models.
Critical strengths
- The proposed method, when extended to constant-curvature space, has small overhead compared to SGW, which is the original method on Euclidean space.
Critical weaknesses
-
Equation 10 (the definition of CCSGW) seems to have substantive issues. In specific:
- The direction measure is inconsistently denoted by and .
- The formula itself doesn't resemble SGW [1] in Equation 5. SGW pads before projection, while CCSGW pads the projection axis before projecting onto it, causing the loss of last dimensions when projecting. Thus, CCSGW wouldn't reduce to Euclidean SGW in [1] as claimed in Section 3.2. For example, with , as projecting onto every padded direction cancels the last coordinate while
-
The mean gains in graph node classification and EEG-based multimodal learning are modest relative to the reported variability. Tables 2 and 3 alone do not establish the statistical reliability of these improvements.
-
Minor:
- Typo "Wegrate" at line 372
Ref:
[1] Vayer et al. Sliced Gromov-Wasserstein. NeurIPS 2019
Questions for authors
-
What is the magnitude of and related to and defined in Theorem 3.1, measured in experiments of Section 4? Since is bounded by , those terms may dominate the bound.
-
Please clarify the motivation of enforcing geometric consistency across components in the mentioned tasks in Section 4. In particular, why should components intended to capture different geometric structures benefit from having their within-component distance structures aligned?