Source-linked AI summary
Geometric deep learning on graphs and manifolds using mixture model CNNs
Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodolà, Jan Svoboda, Michael M. Bronstein
TL;DR
Deep learning has mainly targeted Euclidean data, motivating methods for geometric data such as graphs and manifolds. This paper introduces a spatial framework that unifies prior non-Euclidean CNNs and achieves state-of-the-art results across geometric deep-learning tasks.
Problem
Geometric deep learning needs methods for non-Euclidean data such as graphs and manifolds, but existing spectral approaches are basis-dependent and can be difficult to transfer across domains.
Method
The paper introduces mixture model networks, a spatial-domain framework that uses learnable weighting functions over local pseudo-coordinates and represents prior graph and manifold CNNs as special cases.
Results
The model achieves state-of-the-art results across image, graph, and 3D shape analysis tasks, including significantly outperforming competing approaches on manifold correspondence.
Takeaways & Limitations
MoNet provides a unified spatial framework whose intrinsic construction is deformation-invariant for deformable 3D shape analysis and generalizes across domains.
Takeaways & Limitations
Spectral CNNs remain constrained by graph-specific Fourier bases, while spatial constructions have less clear geometric interpretation on general graphs.
Abstract
from arXiv · showhide
Deep learning has achieved a remarkable performance breakthrough in several fields, most notably in speech recognition, natural language processing, and computer vision. In particular, convolutional neural network (CNN) architectures currently produce state-of-the-art performance on a variety of image analysis tasks such as object detection and recognition. Most of deep learning research has so far focused on dealing with 1D, 2D, or 3D Euclidean-structured data such as acoustic signals, images, or videos. Recently, there has been an increasing interest in geometric deep learning, attempting to generalize deep learning methods to non-Euclidean structured data such as graphs and manifolds, with a variety of applications from the domains of network analysis, computational social science, or computer graphics. In this paper, we propose a unified framework allowing to generalize CNN architectures to non-Euclidean domains (graphs and manifolds) and learn local, stationary, and compositional task-specific features. We show that various non-Euclidean CNN methods previously proposed in the literature can be considered as particular instances of our framework. We test the proposed method on standard tasks from the realms of image-, graph- and 3D shape analysis and show that it consistently outperforms previous approaches.
1. Introduction
Geometric deep learning extends deep learning toward graphs and manifolds, motivated by the growing use of non-Euclidean data across networks, scientific analysis, and computer graphics. The paper introduces a spatial framework whose parametric local patches unify several prior approaches.
- Non-Euclidean data arise in social, transportation, sensor, brain, gene-regulatory, and 3D-shape applications.
- Geometric deep learning seeks to generalize deep learning methods to graph- and manifold-structured data.
- Spectral graph CNNs face domain-dependent Fourier bases, limiting straightforward transfer of models between graphs.
- Spatial methods transfer across domains, but constructions such as anisotropic heat kernels have less clear geometric interpretation on general graphs.
- MoNet defines convolution-like operations through parametric local intrinsic patches represented by mixtures of Gaussian kernels.
- The framework recasts prior manifold and graph CNNs as particular instances and is evaluated on image classification, graph vertex classification, and 3D-shape correspondence.
2. Deep learning on graphs
Graph deep learning extends Fourier-based CNN ideas to graph signals, while addressing domain transfer, computational cost, localization, and parameterization challenges. Several spectral and spatial variants provide different trade-offs in efficiency and locality.
- Spectral graph CNNs: Graph signals use the Laplacian eigenbasis as a generalized Fourier basis, enabling spectral transforms and convolution through element-wise frequency multiplication.The graph Fourier transform is Φ⊤f, and spectral convolution reconstructs the product of transformed signals.
- Spectral graph CNNs: Spectral CNNs apply graph convolutional layers using learnable spectral multipliers on input vertex signals.The framework also includes vertex-wise nonlinearities and graph coarsening as the analogue of pooling.
- Limitations: Spectral CNNs are basis-dependent, costly on general graphs, not guaranteed to be spatially localized, and can require O(n) parameters per layer.Fourier transforms require O(n^2) matrix multiplications when eigenvectors are explicitly used.
- Spectral alternatives: Smooth spectral filters reduce parameter dependence on graph size by interpolating filter coefficients with fixed kernels.The interpolation representation uses a constant number of parameters in the input size n.
- Spectral alternatives: Chebyshev spectral filters avoid explicit eigenvector computation and achieve O(rn) filtering while remaining localized to r-hop neighborhoods.The localization follows from recursively applying the graph Laplacian, whose powers expand the affected neighborhood.
- Spatial graph CNNs: GCN uses a restricted Chebyshev construction with renormalization to stabilize repeated convolutional layers, while DCNN uses powers of the random-walk transition matrix.DCNN produces features from diffusion lengths P^0 through P^(r−1).
3. Deep learning on manifolds
Manifold CNNs represent local surface structure intrinsically through geodesic patches or anisotropic heat kernels. These spatial-domain approaches avoid the cross-domain transfer limitation of spectral methods and support intrinsic shape analysis.
- Manifold foundations: Three-dimensional shapes are modeled as two-dimensional Riemannian manifolds, allowing intrinsic geometric processing of surface data.A Riemannian metric supplies smoothly varying inner products on tangent spaces.
- Unified framework: The proposed framework unifies geometric CNNs by selecting pseudo-coordinates and weight functions for a local patch operator.Table 1 characterizes existing graph and manifold CNNs as particular settings of this construction.
- Geodesic CNN: GCNN maps neighborhood values into geodesic polar coordinates and matches learned templates against the resulting intrinsic patches.Template rotations address the origin ambiguity of the angular coordinate.
- Anisotropic CNN: ACNN uses anisotropic diffusion and heat kernels whose elongation, orientation, and scale vary with α, θ, and t.The conductivity tensor models position- and direction-dependent heat flow on the manifold.
- Comparison: GCNN and ACNN operate spatially, avoiding spectral methods’ inherent inability to generalize across domains and outperforming handcrafted approaches on deformable-shape correspondence.The comparison concerns intrinsic correspondence, described as a difficult computer-graphics problem.
4. Our approach
MoNet provides a generic spatial framework for graph and manifold convolutions, using pseudo-coordinates and learnable weighting functions to construct local patches. Its parametric kernels encompass earlier methods while adding degrees of freedom intended to improve performance.
- Framework: MoNet defines non-Euclidean convolution using pseudo-coordinates and learnable kernel weights over neighborhoods of graph vertices or manifold points.The weighting function is parametrized by learnable parameters, and the resulting patch operator generalizes convolution in the spatial domain.
- Patch construction: The patch operator uses weighting functions over intrinsic local coordinates, with its dimensionality determined by the number of extracted kernel components.On manifolds, the weighting functions can be expressed over local polar coordinates ρ and θ.
- Relations to prior methods: Earlier graph and manifold methods, including GCN, GCNN, and ACNN, arise as particular settings of MoNet's pseudo-coordinates and weight functions.GCNN and ACNN use Gaussian kernels on local polar geodesic coordinates, while GCN uses a triangular kernel based on vertex degree.
- Learnable kernels: Unlike fixed handcrafted weighting functions, MoNet uses parametric kernels with learnable parameters.The proposed construction specifically considers learnable kernels rather than the fixed patches used by previous approaches.
- Learnable kernels: Diagonal Gaussian-mixture covariances give each kernel 2d parameters and the patch operator 2Jd parameters.The covariance matrices are restricted to diagonal form, with Σj and µj serving as learnable covariance and mean parameters.
- Model capacity: The added degrees of freedom provide sufficient architectural complexity for MoNet to outperform existing approaches.The paper also notes that more complex weighting functions could apply nonlinear transformations or network-in-network architectures to the pseudo-coordinates.
5. Results
The proposed MoNet framework is evaluated on image, graph, and 3D-shape tasks, where it generally performs strongly against competing methods. Results highlight robustness to varying graph structures, favorable vertex classification, and superior intrinsic shape correspondence.
- Images: On MNIST, ChebNet performance drops by up to almost 25% on superpixel-based graphs because image-specific graph bases hinder generalization.The degradation is strongest for smaller, more variable graphs.
- Images: MoNet maintains consistently high MNIST accuracy across graph representations, with only light degradation at 75 superpixels.
- Graphs: MoNet compares favorably with DCNN and GCN on vertex classification in the Cora and PubMed citation graphs.The experiments used repeated training and testing to measure average behavior across initializations.
- Manifolds: For dense intrinsic correspondence on FAUST meshes, MoNet significantly outperforms competing approaches, with close to 90% of points at zero error and 99% below 4cm.Correspondence is evaluated by geodesic distance from ground-truth matches.
- Manifolds: MoNet also shows superior correspondence performance to an equivalent Euclidean CNN evaluated on raw depth values and SHOT descriptors.
6. Conclusions
The paper proposes a spatial-domain model for deep learning on graphs and manifolds and presents it as a generalization of several previous techniques. Experiments across geometric learning tasks report state-of-the-art results, while 3D-shape applications benefit from intrinsic deformation invariance.
- MoNet is a spatial-domain model for deep learning on non-Euclidean domains including graphs and manifolds.
- Several previous geometric deep learning techniques can be obtained as particular instances of the proposed framework.
- Extensive experiments show applicability across different geometric deep learning tasks and state-of-the-art results.
- For deformable 3D-shape analysis, the intrinsic formulation is deformation-invariant by construction, unlike Euclidean models that generally require greater complexity and much larger training sets to learn deformation invariance.