Learning Topology-Driven Multi-Subspace Fusion for Grassmannian Deep Network

AI summary

Overview

Research area: Geometric deep learning on Riemannian manifolds, specifically Grassmannian manifold representation learning for computer vision, EEG signal classification, and graph learning.

Technical level: Advanced. The paper assumes familiarity with Riemannian geometry, Grassmannian manifolds, tangent spaces, exponential/logarithmic maps, Fréchet means, and topological convergence arguments.

Scope: The paper proposes GMSF-Net, a topology-driven framework that replaces the static single-subspace assumption of prior Grassmannian networks with adaptively constructed and fused multiple subspaces.

What This Paper Is About

Existing Grassmannian deep networks represent high-dimensional data using a single, static low-dimensional subspace, which the authors argue cannot capture local geometric variation or multi-modal data distributions. The paper's goal is to build a network that learns several subspaces at once, interacts them, and fuses them into one representation, while proving that this construction converges stably. The authors frame this as transferring the multi-channel interaction idea of Euclidean networks (for example, LeNet-5 style stacked convolution blocks) into the non-Euclidean Grassmannian domain.

Key Contributions

1. Adaptive multi-subspace modelling. A mechanism (AdaMSC) that dynamically selects and weights task-relevant subspaces, inspired by the Kolmogorov-Arnold representation theorem, with subspace selection driven by topological convergence analysis.
2. Grassmannian Multi-Subspace Fusion network (GMSF-Net). A network that fuses heterogeneous subspace representations, including a Multi-Subspace Interaction Block that combines Grassmannian Multi-Subspace Representations (GMSR) with Grassmannian Subspace Interaction (GSI) via Fréchet mean optimisation.
3. A topology-driven framework with theoretical guarantees. Convergence of adaptive Grassmannian subspaces is established under a projection metric topology, so that iteratively updated subspaces converge to a stable subspace S* and gradient-based optimisation remains stable.
4. Practical regularisation and stacking. Riemannian batch normalisation (SPDBN) and a mutual-information-inspired regulariser are integrated into a stackable architecture to improve discriminability, robustness, and feature-space topology stability.

Main Findings

• FPHA 3D action recognition (Table 1): GMSF-Net-3Blocks reaches 91.22% ± 0.53, versus GMSF-Net-2Blocks at 90.70% ± 0.70 and GMSF-Net-1Block at 90.43% ± 0.74. Comparison methods: SPDNetBN 89.33% ± 0.49, MATT 87.70% ± 0.68, SPDNet 87.65% ± 1.02, GDLNet 87.60% ± 0.69, GrNet 78.79% ± 1.82. The paper reports a 12.43% improvement over GrNet.
• HDM05 3D action recognition (Table 2): GMSF-Net-3Blocks reaches 64.19% ± 0.88, with 2Blocks at 63.98% ± 1.07 and 1Block at 63.64% ± 1.24. Comparison methods: MATT 62.25% ± 1.68, SPDNet 60.45% ± 1.12, GDLNet 60.08% ± 1.78, GrNet 59.23% ± 1.78. The reported improvement over GrNet is up to 4.96%.
• MAMEM-SSVEP-II EEG classification (Table 3): GMSF-Net-3Blocks reaches 66.87% ± 1.46, with 1Block at 66.74% ± 1.79 and 2Blocks at 66.32% ± 1.84. Comparison methods include GDLNet 65.52% ± 2.86, MATT 65.19% ± 3.14, SPDNetBN 62.76% ± 3.01, SPDNet 62.30% ± 3.12, SCCNet 62.11% ± 7.70, GrNet 61.23% ± 3.56, ShallowConvNet 56.93% ± 6.97, MBEEGSE 56.45% ± 7.27, EEG-TCNet 55.45% ± 7.66, EEGNet 53.72% ± 7.23, FBCNet 53.09% ± 5.67. Reported gains: 5.64% over GrNet, 1.68% over MATT, and 1.35% over GDLNet, with improved stability.
• Efficiency advantage: On FPHA, GMSF-Net-3Blocks uses 1.30 MB and 81.07M FLOPs versus GrNet's 6.73 MB and 38.60M FLOPs and SPDNetBN's 13.63 MB and 1902.97M FLOPs; on HDM05 it uses 3.39 MB and 114.12M FLOPs versus SPDNet's 11.71 MB and 2050.49M FLOPs.
• Graph tasks (Table 4): GMSF-Net-3Blocks achieves Disease LP 95.5% ± 0.4 / NC 82.7% ± 0.8, Airport LP 91.2% ± 0.2 / NC 78.3% ± 0.6, PubMed LP 93.6% ± 0.3 / NC 72.9% ± 0.3, and Cora LP 88.9% ± 0.3 / NC 56.2% ± 1.1. The paper states gains hold on all graph datasets except PubMed, and that gains are larger on datasets with clearer subspace structure (Disease, Cora) than on structurally complex datasets deviating from ideal subspace assumptions (PubMed, Airport).
• Ablation on subspace construction and interaction (Table 5): The combination of adaptive subspace plus interaction is best on all three datasets (HDM05 63.64% ± 1.24, FPHA 90.43% ± 0.74, SSVEP 66.74% ± 1.79). Random subspace is weakest with interaction (HDM05 50.29% ± 2.10, FPHA 72.47% ± 3.19, SSVEP 56.01% ± 2.46). Fixed subspace with interaction reaches 53.04% ± 0.88 on HDM05, 83.06% ± 0.61 on FPHA, and 66.05% ± 1.95 on SSVEP.
• Random subspace can degrade performance: The paper notes HDM05 dropping from 52.50% (random subspace, no interaction) to 50.29% (random subspace, with interaction), indicating low-quality interactions may introduce noise.
• Performance saturation with depth: Increasing the number of interaction blocks from one to three yields diminishing returns on all reported tasks.
• Convergence result: Under the projection metric topology, the subspace sequence satisfies d(S'(t), S*) → 0 as t → ∞, meaning iterates converge to a stable subspace.

Methodology in Plain English

The input data (for example, a video clip or an EEG trial) is first summarised as a covariance matrix that captures statistical dependencies, and this matrix is decomposed via Schmidt orthogonalisation into a set of atomic one-dimensional subspaces. Rather than trusting any single subspace, the network keeps a set of atomic subspaces and learns, through Softmax-normalised importance weights, which of them to pick and how heavily to weight them when building each of several new, richer subspaces. Only the top-p atoms are used per new subspace, which keeps the representation compact and task-adaptive.

Once these candidate subspaces exist, a second stage applies several learnable mapping matrices (constrained to the Stiefel manifold so the geometry is preserved) to produce different "views" of each subspace under different geometric frameworks. These views are then merged. When only two subspaces are merged, the fusion has a closed-form solution that lies exactly on the geodesic between them; for more than two, the Karcher flow algorithm iteratively walks through tangent spaces until it reaches the Fréchet mean, the point minimising the weighted sum of squared geodesic distances. This gives a single fused subspace that respects manifold geometry instead of flattening it.

The whole pipeline is made deep and stackable by inserting Riemannian batch normalisation, which maps statistical features onto the SPD manifold and normalises them so the topology of the feature space stays stable across layers. Training uses cross-entropy loss plus a mutual-information-inspired regulariser weighted by a hyperparameter λ, encouraging the subspaces to be complementary rather than redundant. Because updates must live on the manifold, optimisation uses Riemannian gradient descent with tangent-space projection and the exponential map; the authors analyse this update through a projection-metric topology to argue that it converges.

Why This Matters

• Research impact: The paper challenges the static single-subspace convention that has dominated Grassmannian deep learning since GrNet, and claims to be the first work introducing deep interactions of Grassmannian subspaces in Riemannian neural networks. It supplies convergence guarantees rather than relying on empirical success alone, and it carries the multi-channel interaction philosophy of Euclidean CNNs into non-Euclidean geometry.
• Real-world applications:
  • 3D human action recognition from video (evaluated on FPHA and HDM05).
  • Brain-computer interfaces and EEG-based classification (evaluated on MAMEM-SSVEP-II).
  • Graph-structured prediction such as node classification and link prediction (evaluated on Disease, Airport, PubMed, Cora).
  • Other tasks the introduction cites for Grassmannian modelling: emotion recognition and face verification.
• Industry relevance: The reported model sizes (1.30 MB on FPHA, 3.39 MB on HDM05, 1.94 MB on SSVEP) and reduced FLOPs relative to SPD-manifold baselines suggest deployment on resource-constrained hardware, which matters for wearable EEG devices, embedded vision, and edge robotics. The public code release at https://github.com/Xua-Yu/GMSF-Net lowers the barrier to adoption.

Future Directions

• Extending to domains where the Grassmannian assumption is weaker. The paper itself observes that performance gains shrink on PubMed and Airport, which deviate from ideal subspace structure; a natural next step is handling data far from the manifold assumption.
• Scaling depth beyond three blocks. Since performance saturates at three blocks, it is an open question whether architectural changes could make deeper subspace stacks continue to help.
• Broader modality transfer. The conclusion names computer vision and natural language processing as target fields for this paradigm, but no NLP experiments are reported in this content.
• Unspecified hyperparameter and dataset details. The balance hyperparameter λ and the datasets' sizes are not reported in the provided content, so sensitivity analysis and scaling behaviour remain open questions here.

Target Audience

Researchers in geometric and Riemannian deep learning, graduate students working on manifold-valued representation learning, and practitioners in 3D action recognition, EEG-based brain-computer interfaces, and graph learning who need compact models with theoretical convergence support. Readers without a background in differential geometry and topology will find the theoretical sections difficult, though the architectural and experimental sections are more accessible.

Authors’ abstract

Grassmannian manifold offers a powerful carrier for geometric representation learning by modelling high-dimensional data as low-dimensional subspaces. However, existing approaches predominantly rely on static single-subspace representations, neglecting the dynamic interplay between multiple subspaces critical for capturing complex geometric structures. To address this limitation, we propose a topology-driven multi-subspace fusion framework that enables adaptive subspace collaboration on the Grassmannian. Our solution introduces two key innovations: (1) Inspired by the Kolmogorov-Arnold representation theorem, an adaptive multi-subspace modelling mechanism is proposed that dynamically selects and weights task-relevant subspaces via topological convergence analysis, and (2) a multi-subspace interaction block that fuses heterogeneous geometric representations through Fréchet mean optimisation on the manifold. Theoretically, we establish the convergence guarantees of adaptive subspaces under a projection metric topology, ensuring stable gradient-based optimisation. Practically, we integrate Riemannian batch normalisation and mutual information regularisation to enhance discriminability and robustness. Extensive experiments on 3D action recognition (HDM05, FPHA), EEG classification (MAMEM-SSVEPII), and graph tasks demonstrate state-of-the-art performance. Our work not only advances geometric deep learning but also successfully adapts the proven multi-channel interaction philosophy of Euclidean networks to non-Euclidean domains, achieving superior discriminability and interpretability.