Spherical CNNs
For computer vision and the natural sciences problems that require the analysis of spherical data, where representations may be learned efficiently by encoding equivariance to rotational symmetries
# Resources
# Code
- #CODE
DeepSphere
- Learning on the sphere (with a graph-based ConvNet). Used so far for cosmology, geophysics, 3D object recognition.
# References
- #PAPER Spherical CNNs (Cohen 2018)
- #PAPER Learning SO(3) Equivariant Representations with Spherical CNNs (Esteves 2018)
- #PAPER Spherical CNNs on unstructured grids (Jiang 2019)
- #PAPER
Spherical convolution and other forms of informed machine learning for deep neural network based weather forecasts (Scher 2020)
- CNN-based weather forecasting solutions are are usually trained on atmospheric data represented as regular latitude-longitude grids, neglecting the curvature of the Earth
- Showed the benefit of replacing the convolution operations with a spherical convolution operation, which takes into account the geometry of the underlying data, including correct representations near the poles
- Additionally, studied the effect of including the information that the two hemispheres of the Earth have “flipped” properties - for example cyclones circulating in opposite directions - into the structure of the network
- Using spherical convolution leads to an additional improvement in forecast skill, especially close to the poles in the first days of the forecast
- The spherical convolution is implemented flexibly and scales well to high resolution datasets, but is still significantly more expensive than a standard convolution operation
- #PAPER DeepSphere: a graph-based spherical CNN (Defferrard 2020)
- #PAPER Efficient Generalized Spherical CNNs (Cobb 2021)