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 



Full paperNeural Feature Learning in Function Space. (JMLR, vol 25:142)

More information can be found in this blog, including some pytorch demos.