acceptodds
Withdrawn from ICLR 2027

Learning Distributions of Grassmann Subspace Combinations for Effective Geometric Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Grassmann manifolds provide a principled framework for modelling data as low-dimensional subspaces, capturing intrinsic geometric structure and achieving considerable success in geometric representation learning. Existing approaches, however, typically treat all inner subspace combinations within a Grassmann point as equally important, overlooking their non-uniform distribution. To address this issue, we theoretically characterise the distribution of subspace combinations within a Grassmann point based on spectral measure theory and quadratic utility theory, revealing that different combinations contribute non-uniformly to representation. This insight suggests that Grassmann representation learning should adaptively model such distributions and exploit compatibility among low-dimensional subspaces to select and construct effective combinations. To this end, we propose a **Gra**ph-**S**tructured adaptive **S**ubSpace (GraSS) modelling on Grassmann manifolds. GraSS builds a relational graph among constituent subspaces to jointly capture their importance and structural compatibility. Based on the constructed graph, we identify the maximum connected subgraph with high coherence and accordingly adaptively determine the mutually compatible constituent subspaces and their composition. Furthermore, we introduce an adaptive gating selection mechanism to dynamically filter the less powerful combinations, suppressing redundant combinations while emphasising those with higher representation utility. Finally, we apply GraSS to multiple downstream tasks. Extensive experiments show that GraSS consistently improves existing Grassmann representation learning methods across tasks and exhibits strong generalisation.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Grassmann Manifold, Graph-Structured Modelling, Subspace Representation Learning

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

This paper proposes GraSS, which uses a graph to learn the importance and mutual compatibility of constituent subspaces. It constructs Grassmann representations by combining subspaces within each connected component and applying gating-based selection. The authors present a theory based on spectral measures and quadratic utility and evaluate the proposed method on action recognition using HDM05, FPHA, and NTU60 with 10 random seeds. In addition to comparisons with existing methods based on Riemannian geometry and component ablations, they analyze the relationship between the complexity of individual samples and the required dimensionality on Salinas.

Critical strengths
  • Table 1 shows that, under the reported NTU60 setting, GraSS improves accuracy from GMSF-Net’s 58.52% to 61.04%, while reducing the standard deviation from 0.99 to 0.29. These results show empirical improvements in both performance and stability across runs.
  • Table 2 shows step-by-step improvements on FPHA: 81.99% with fixed partitioning, 88.08% with the graph, and 90.44% with the full model. These results suggest that learning a relational structure and introducing a selection mechanism provide benefits beyond simply partitioning subspaces.
Critical weaknesses
  • If Eq. (21) takes a weighted sum of the basis vectors of one-dimensional subspaces as written, wouldn’t the rank remain at most 1 even after QR decomposition? In that case, the dimension of the composite subspace would not increase. Also, flipping the sign of a basis vector may change the composite subspace. This may indicate an inconsistency with the central claims of adaptive dimensionality selection and basis invariance.
  • The measure in Proposition 1 depends on the singular values of the original matrix and is not uniquely determined by the Grassmann point alone. The paper also does not show why graph partitioning or gating-based selection satisfies the assumptions of Theorem 3. The connection between the theory of an “intrinsic distribution” and the proposed algorithm is not sufficiently established.
Questions for authors
  • Does Eq. (21) take a weighted sum of the basis vectors? Please provide the tensor shapes and explain how the dimension of the composite subspace can vary and how invariance to sign flips of singular vectors is ensured.
  • Beyond the initialization in Eq. (19), what input information is used to learn compatibility? Please also clarify how the presence or absence of edges is determined and describe the training procedure, including connected-component extraction.
  • Does Proposition 1 describe a distribution associated with the original matrix, or a distribution intrinsic to the Grassmann point? Please also explain whether there is a guarantee that the actual selection results satisfy the assumptions of Theorem 3.
Reviewer_N1HFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a graph based method that learns the relationships among constituent subspaces to guide their combination, followed by a selection mechanism (gating) to obtain more effective Grassmannian representations. The key contribution is the use of a learned relational graph to determine how constituent subspaces should be grouped and combined based on their compatibility. The adaptive gating mechanism further suppresses incompatible or redundant subspaces.

Critical strengths
  • The paper identifies a key limitation in current Grassmannian networks: they treat constituent subspaces equally or select them based on individual importance, without explicitly accounting for their compatibility. Considering subspace compatibility and relationships is intuitive and well-motivated.
Critical weaknesses
  • The reported experimental results contain unexplained inconsistencies, which reduce confidence in the reported improvements and make it difficult to assess whether the evidence supports the paper’s main claims.
  • The mathematical formulation does not clearly align with the method description and Figure 2, leaving the core algorithm insufficiently specified and making its technical correctness difficult to assess.
Questions for authors
  • Could the authors clarify why the GraSS results reported in Table 2, 63.72% (HDM05) and 90.44% (FPHA), do not align with the values in lines 429–430, 63.12% and 90.34%, respectively?
  • In Eq. 21, if Qi∈RM×1Q_i \in \mathbb{R}^{M \times 1} is a basis vector, the proposed weighted composite subspace representation, defined by Q~=∑i∈CkaikQi\tilde{Q} = \sum_{i \in C_k} a_{ik}Q_i, remains a single vector, and the subsequent QR decomposition cannot produce a higher-dimensional subspace. However, in Figure 2, the features look like they are concatenated. Is the intended operation a weighted sum or weighted concatenation? Please specify the resulting dimensions and clarify how this operation implements the claimed adaptive subspace dimensionality.
  • Table 1 reports exactly the same FLOPs (0.6953M) for GraSS and GrNet, despite GraSS introducing additional graph modelling, connected-component analysis, subspace fusion, and gating operations. Is this because these operations are considered lightweight and therefore excluded from the FLOPs calculation? Could the authors clarify how FLOPs were calculated and whether these additional operations were included in the reported computational cost?
Reviewer_j39zOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

GraSS learns a graph over one-dimensional SVD subspaces, fuses bases within connected components, and applies gates before a Grassmann network. The stated goal is to adapt constituent grouping and selection without manually fixing the number and dimension of composite subspaces. A quadratic-utility analysis separates individual importance from pairwise compatibility and gives a conditional omission bound. Experiments cover HDM05, FPHA, the mutual-action subset of NTU60, component ablations, and a Salinas dimensionality analysis.

Critical strengths
  1. Importance versus compatibility. Theorem 1 and Proposition 2 make explicit why individually useful directions can interact unfavourably under the specified utility, providing a clear motivation for studying combinations.
  2. Repeated runs and ablations. Section 6.1 reports ten seeds; Table 2 compares composition, graph, and gating variants. Table 1 shows a 2.52-point gain over GMSF-Net on the evaluated NTU60 setting, with comparable FPHA/HDM05 performance.
Critical weaknesses
  1. Fusion lacks the claimed geometric properties. Section 3.1 and Appendix D.1 require basis invariance, but Eq. (21) sums specific one-dimensional bases. For two equally weighted orthogonal vectors, reversing just one vector leaves the input subspaces unchanged while changing the line spanned by their sum; QR in Eq. (22) cannot fix this. Each basis also has one column, so every nonzero fused component remains rank one. Changing the number of retained lines does not establish the variable-rank compositions depicted in Fig. 1. Appendix E.1 proves only constituent-projector sign invariance. Please define the intended object and a basis-invariant composition with explicit output rank, then verify these properties in the implementation.
  2. Adaptive selection and its theoretical justification remain unclear. Because Section 5.1 uses mutually orthogonal vectors from one SVD, Eq. (2) gives projection distance one between every distinct pair; Eq. (19) therefore initializes identical off-diagonal weights. The learned adjacency must supply adaptivity, yet Sections 5.2-5.3 and Algorithm 1 leave its sample dependence, discrete edge/gate rules, and gradients through connected components unspecified. Appendix F does provide L1 graph and continuous sigmoid-gate penalties, but these alone do not define discrete selections; Appendix G.2 adds no such rule. Theorem 3's marginal-gain conditions are not mapped to the algorithm. Table 2 demonstrates predictive gains, while Salinas establishes heterogeneous dimensionality needs; neither verifies actual adaptive selections. Please specify the executable selection/training rule, report sample-wise component counts and ranks, and connect the rule to the bound or narrow the claim.
Questions for authors
  1. What basis-invariant geometric object, with what rank, is actually returned by Eqs. (21)-(22)?
  2. How does the implemented graph and gate rule produce sample-dependent discrete selections from the orthogonal SVD directions?

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

Related papers

Discussion (0)

Sign in to comment.