Abstract
Group convolutional neural networks (G-CNNs) are powerful tools for exploiting known symmetries in signals. However, they typically require signals to be defined directly on the group, often necessitating a computationally expensive "lifting" step. We propose a novel convolutional filter derived directly from the Lie group algebra, bypassing the need for lifting. By leveraging Algebraic Signal Processing (ASP) and establishing a connection to multigraph signal processing, we ensure both computational efficiency and mathematical stability. Our approach is validated on $SO(3)$ symmetry classification tasks, demonstrating superior performance and scalability compared to existing methods.
Introduction: The Challenge of Group Convolutions
Symmetry is a fundamental inductive bias in deep learning, with standard CNNs exploiting translation invariance. G-CNNs generalize this to more complex symmetries like rotations ($SO(3)$). A significant bottleneck in current G-CNN architectures is the requirement that signals reside on the group manifold. When data exists in other domains (e.g., 3D point clouds), it must be "lifted" to the group, a process that is often heuristic-heavy, memory-intensive, and difficult to scale.
Methodology: An Algebraic Approach
To address these limitations, we utilize Algebraic Signal Processing (ASP). In this framework, a signal model is defined by an algebra of filters, a vector space of signals, and a homomorphism describing their interaction.
Our core innovation involves extending ASP to continuous Lie groups using Banach *-algebras. This allows us to define filters as elements of the Lie group algebra rather than functions on the group itself. To make this continuous framework computationally tractable, we implement a two-step relaxation:
- Group Discretization: We approximate the group using the Lie algebra and an exponential map.
- Signal Interpolation: We use barycentric (natural neighbor) interpolation to handle the induced group action on sampled signal spaces.
Crucially, we prove that this discretized filter is equivalent to a filter on a multigraph, allowing us to inherit formal stability guarantees against deformations and approximation errors.
Experimental Evaluation
We evaluated our "GrpA" (Group Algebra) filters on two tasks with $SO(3)$ rotational symmetry:
- Knot Classification: Distinguishing between Trefoil and Listing’s knots under various sampling schemes (Grid, Uniform, Sphere).
- ModelNet10: A benchmark 3D CAD model dataset.
Key Results
Our experiments highlight several advantages of the algebraic approach:
- Scalability: Unlike baselines like LieConv, which become intractable as point cloud density increases (e.g., at 1,000 samples), GrpA scales efficiently due to its avoidance of the lifting step.
- Accuracy: The two-layer GrpA-2 model consistently outperformed standard Fully Connected Networks (FCNN) and LieConv across different sampling geometries.
- Offline Efficiency: The matrices representing group actions can be precomputed, significantly accelerating training and inference.
- Geometric Sensitivity: Results indicate that the choice of sampling (e.g., Sphere vs. Grid) significantly impacts performance, with spherical sampling generally proving most effective for rotational tasks.
Conclusion and Future Directions
By defining convolutional filters within the Lie group algebra, we provide a mathematically rigorous and computationally efficient alternative to traditional lifting-based G-CNNs. While we have demonstrated empirical stability and performance, future work will focus on establishing formal error bounds for the discretization process and optimizing sampling strategies through signal bandwidth analysis on Lie group algebras.
Summary updated