Abstract
Group equivariant convolutional networks (GCNNs) endow classical convolutional networks with additional symmetry priors, which can lead to a considerably improved performance. Recent advances in the theoretical description of GCNNs revealed that such models can generally be understood as performing convolutions with G-steerable kernels, that is, kernels that satisfy an equivariance constraint themselves. While the G-steerability constraint has been derived, it has to date only been solved for specific use cases - a general characterization of G-steerable kernel spaces is still missing. This work provides such a characterization for the practically relevant case of G being any compact group. Our investigation is motivated by a striking analogy between the constraints underlying steerable kernels on the one hand and spherical tensor operators from quantum mechanics on the other hand. By generalizing the famous Wigner-Eckart theorem for spherical tensor operators, we prove that steerable kernel spaces are fully understood and parameterized in terms of 1) generalized reduced matrix elements, 2) Clebsch-Gordan coefficients, and 3) harmonic basis functions on homogeneous spaces.
Introduction
In deep learning, incorporating symmetry priors is a powerful way to improve model performance and generalization. Group Equivariant Convolutional Networks (GCNNs) achieve this by ensuring that if an input is transformed by a symmetry operation (like a rotation), the output transforms predictably. This property, known as equivariance, is enforced by using "steerable" kernels—convolution filters that satisfy a specific mathematical constraint to ensure they behave predictably under symmetry transformations.
Historically, solving this steerability constraint has been a manual, group-specific process. Researchers had to derive unique solutions for every specific group, such as 2D rotations ($SO(2)$) or 3D rotations ($SO(3)$). This paper addresses the lack of a general theory by providing a unified characterization of steerable kernel spaces for any compact symmetry group, effectively turning an ad-hoc derivation process into a rigorous mathematical framework.
Methodology: The Physics-AI Connection
The core insight of this work is a striking parallel between deep learning and quantum mechanics. The constraints on convolution kernels in GCNNs are mathematically identical to those on "spherical tensor operators" in physics. By generalizing the Wigner-Eckart theorem—a cornerstone of quantum mechanics used to simplify matrix element calculations—the authors provide a systematic way to parameterize equivariant kernels.
The derivation involves decomposing the space of steerable kernels through a series of isomorphisms, moving from the abstract space of steerable kernels to a concrete parameterization based on group representation theory.
This decomposition reveals three fundamental building blocks:
- Harmonic Basis Functions: Derived from the Peter-Weyl theorem, these functions (e.g., spherical harmonics) provide a natural basis for the space.
- Clebsch-Gordan Coefficients: Algebraic coupling coefficients that describe how different group representations interact.
- Endomorphisms: Representing the "reduced matrix elements" from physics, these are the actual learnable parameters in the network.
While the harmonics and Clebsch-Gordan coefficients are fixed by the geometry of the symmetry group, the endomorphisms represent the degrees of freedom that the optimizer adjusts during training.
Key Results and Insights
The primary contribution is the proof of Theorem 4.1, which provides a complete characterization of steerable kernel spaces for any compact group. This framework was validated by reproducing existing, manually-derived kernels for groups like $SO(2)$, $SO(3)$, and $\mathbb{Z}/2$. Furthermore, it enabled the derivation of new kernel bases for more complex groups, such as $O(3)$ (rotations and reflections in 3D), which were previously uncharacterized.
The paper also clarifies the role of real versus complex representations. In complex-valued networks, Schur’s Lemma often reduces the learnable space to a single scaling factor. However, in real-valued networks, the space of endomorphisms can be 1, 2, or 4-dimensional, explaining the inherent complexity in designing real-valued equivariant models.
Conclusion and Future Work
This work bridges the gap between theoretical physics and machine learning, providing a standardized procedure for designing equivariant layers. However, the current theorem is limited to compact groups due to its reliance on the Peter-Weyl theorem. Future research may extend this to non-compact groups, such as scaling or the Lorentz group, using techniques like Pontryagin duality. Additionally, while the mathematical foundation is now solid, the computational efficiency of calculating Clebsch-Gordan coefficients for large groups remains a practical hurdle for implementation.
Summary updated