Feature Geometry
A mathematical framework for feature-centric information processing, which
- unfies feature learning and information decomposition
- separates feature learning and feature usages
- provides a principled algorithm design
- for learning multivariate dependence structures
- with deep neural networks as building blocks
Applications
- multiuser detection in wireless communication
- sequential dependence decomosition
- understanding kernel methods
Full paper: A Geometric Framework for Neural Feature Learning.
More information can be found in this 
