Feature Geometry
A mathematical framework for feature-centric information processing, which
- formulates representation learning as information decomposition
- separates feature learning and feature usage
- provides principled deep-learning designs for
- adapting learned features
- learning multivariate dependence structures
- computing information measures
Applications
- separable computation of information measures
- learning-based receiver design for wireless communication (MILCOM 2024)
- separable design for feature regularization & adaptation (Allerton 2024)
- learning-based operator SVD solver (ICML 2024)
- feature learning for decomposing sequential dependence (Allerton 2023)
- understanding kernel methods, quantifying kernel quality (ISIT 2023)
More Details: Neural Feature Learning in Function Space. (JMLR, vol 25:142)
This 