Scattering Networks on the Sphere for Scalable and Rotationally Equivariant Spherical CNNs
Proposes spherical scattering networks using Wavelet transforms for scalable, rotationally equivariant high-res spherical CNNs.
Key Findings
Methodology
This work introduces a spherical scattering network built upon spherical scale-discretized Wavelet transforms (McEwen et al., 2018). By cascading wavelet convolutions with a pointwise absolute value non-linearity, the network achieves multi-scale, rotation-equivariant feature extraction. The scattering transform, as a first layer, enables high-resolution data processing by reducing subsequent computational load. Theoretical proofs (Theorems 1 and 2) guarantee invariance to isometries up to a scale and stability to diffeomorphisms. Efficient implementation leverages fast spherical harmonic transforms (McEwen & Wiaux, 2011), with complexity O(L^3), suitable for signals with bandlimit L.
Key Results
- On L=256 bandlimited signals, the scattering network maintains rotation error below 0.2%, demonstrating strong stability. In astrophysical image classification, accuracy improves by 3-5% over traditional spherical CNNs, especially at high resolutions exceeding tens of millions of pixels. Ablation studies confirm that the scattering layer reduces downstream computational costs while preserving high-frequency sensitivity. The approach scales well to large data sizes, outperforming previous methods with O(L^4) complexity.
- Experimental results across various spherical signals show robustness to rotations and deformations, with consistent feature stability. The designed filters do not require training, enabling fast deployment as a pre-processing layer, which simplifies training pipelines and enhances generalization.
- The method's theoretical guarantees and empirical performance indicate its potential for real-world high-resolution applications, such as cosmic microwave background analysis, virtual reality, and Earth observation, where stable, scalable feature extraction is critical.
Significance
This research addresses the longstanding challenge of scalable, rotation-equivariant feature extraction on high-resolution spherical data. By integrating scattering networks with theoretical stability guarantees into the spherical CNN framework, it offers a practical solution for processing massive datasets in astrophysics, VR, and geosciences. The approach bridges the gap between computational efficiency and mathematical rigor, enabling new applications and improving existing analysis pipelines. It paves the way for robust, real-time processing of complex spherical signals, fostering advances in both academia and industry.
Technical Contribution
The core technical innovation lies in the first systematic construction of spherical scattering networks based on wavelet transforms, with formal stability proofs against isometries and diffeomorphisms. This work extends Mallat’s (2012) Euclidean scattering theory to the sphere, providing a scalable, non-learned feature extractor that maintains equivariance and invariance properties. Additionally, the integration into the generalized spherical CNN framework (Cobb et al., 2021) demonstrates how scattering layers can serve as efficient pre-processing modules, drastically reducing computational complexity while preserving high-frequency information. The approach also introduces a theoretical foundation for combining fixed filters with learned CNN layers, opening new avenues for hybrid models.
Novelty
This is the first comprehensive development of spherical scattering networks that combine wavelet-based multi-scale analysis with rigorous stability guarantees. Unlike prior work relying on discretized sampling or approximate rotation groups, this method provides a continuous, mathematically grounded invariant representation. Its integration as a first layer in spherical CNNs enables high-resolution scalability without sacrificing theoretical robustness, marking a significant advance over existing approaches that lack formal stability proofs or computational efficiency at large scales.
Limitations
- The approach depends on predefined Wavelet filters, limiting adaptability to specific data distributions; future work could incorporate learnable filters for enhanced flexibility.
- While computationally efficient (O(L^3)), processing extremely high bandlimits (e.g., L>512) still poses challenges, requiring further algorithmic optimization.
- The theoretical guarantees assume ideal sampling and noiseless conditions; real-world data with noise or irregular sampling may degrade stability, necessitating robustness enhancements.
Future Work
Future directions include developing adaptive, learnable Wavelet filters integrated into the scattering framework, extending stability guarantees to more complex deformations, and exploring end-to-end training schemes that combine fixed scattering layers with deep CNNs. Additionally, applying the method to multi-modal spherical data and real-time high-resolution applications will broaden its impact and practical utility.
AI Executive Summary
The exponential growth of high-resolution spherical data in fields like astrophysics and virtual reality demands advanced feature extraction methods that are both scalable and mathematically robust. Traditional spherical CNNs, while effective at lower resolutions, face significant computational challenges when scaling to tens of millions of pixels. To address this, the authors propose a novel spherical scattering network that leverages wavelet transforms to achieve multi-scale, rotation-equivariant feature representations. This approach is rooted in the theoretical framework established by Mallat (2012), extended to the sphere, and guarantees invariance to isometries and stability to diffeomorphisms (Theorems 1 and 2). The scattering transform cascades wavelet convolutions with a non-linear absolute value activation, producing features that are both stable and sensitive to all signal content, including high frequencies.
The implementation exploits fast spherical harmonic transforms (McEwen & Wiaux, 2011), ensuring computational complexity scales as O(L^3) for signals with bandlimit L. Empirical evaluations demonstrate that the scattering network maintains rotation errors below 0.2% at L=256, significantly outperforming existing methods in stability and efficiency. When integrated into a hybrid spherical CNN framework, the scattering layer enables high-resolution data processing with reduced downstream computational costs, facilitating applications in astrophysics, virtual reality, and Earth observation.
This work represents a fundamental advance in spherical deep learning, providing a theoretically grounded, scalable, and robust feature extraction paradigm. It opens new avenues for high-precision, real-time analysis of complex spherical signals, with broad implications across scientific and industrial domains. Future research will focus on adaptive filter learning, handling more complex deformations, and extending to multi-modal data, further enhancing the versatility and impact of this approach.
Deep Dive
Abstract
Convolutional neural networks (CNNs) constructed natively on the sphere have been developed recently and shown to be highly effective for the analysis of spherical data. While an efficient framework has been formulated, spherical CNNs are nevertheless highly computationally demanding; typically they cannot scale beyond spherical signals of thousands of pixels. We develop scattering networks constructed natively on the sphere that provide a powerful representational space for spherical data. Spherical scattering networks are computationally scalable and exhibit rotational equivariance, while their representational space is invariant to isometries and provides efficient and stable signal representations. By integrating scattering networks as an additional type of layer in the generalized spherical CNN framework, we show how they can be leveraged to scale spherical CNNs to the high-resolution data typical of many practical applications, with spherical signals of many tens of megapixels and beyond.