Core Concepts
Efficiently compute directional wavelet transforms on spherical domains for data analysis.
Abstract
The article discusses the development of new directional wavelet transforms on the sphere and ball, providing up to a 300-fold acceleration compared to existing software. These transforms are automatically differentiable, facilitating integration with modern machine learning techniques. The algorithms are designed for high-performance deep learning research, supporting efficient automatic differentiation and distribution over hardware accelerators. The article also highlights the importance of gradient information for back-propagation during model training and the benefits of high throughput evaluation on hardware accelerators.
Stats
Up to a 300-fold acceleration for signals on the sphere compared to existing software.
Up to a 21800-fold acceleration for signals on the ball compared to existing software.
Quotes
"We publicly release both S2WAV and S2BALL , open-sourced JAX libraries for our transforms that are automatically differentiable."
"These algorithms dramatically accelerate existing spherical wavelet transforms."
"The gradient information afforded by automatic differentiation unlocks many data-driven analysis techniques previously not possible."