Scaling Spherical Deep Learning for High-Resolution Input/Output Data
We recently published another two articles in Towards Data Science (TDS) to provide a more accessible entry into our research.
In a previous post we discussed how spherical CNNs can be scaled to support high-resolution input data. However, such an approach cannot support high-resolution output data, which is necessary for dense prediction tasks, such as semantic segmentation and depth estimation.
In a series of two TDS articles we discuss how we solve the open problem of scaling geometric AI techniques for spherical data to support high-resolution input and output data.
In the first TDS article, titled Geometric Deep Learning on Groups, we review the dichotomy between continuous and discrete approaches. In the second TDS article, titled Hybrid Discrete-Continuous Geometric Deep Learning, we describe how to break this dichotomy through a hybrid discrete-continuous (DISCO) method that provides both excellent performance and high computational scalability.
Our underlying geometric AI technology has now reached a maturity where it can be applied to realise the next wave of open generative AI for geometries like the sphere and 3D more generally. You can learn more on our dedicated CopernicAI website.