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)