Research

Feature Geometry

Feature geometry is a mathematical framework for feature-centric information processing.

• It formulates representation learning as information decomposition.
• It separates feature learning and feature usage.
• It provides principled deep-learning designs for adapting learned features, learning multivariate dependence structures, and computing information measures.

The framework is developed in Neural Feature Learning in Function Space, published in the Journal of Machine Learning Research, 25(142), 2024.

• View the Framework Illustration
• Code and Experiments

Applications

• Information Measures. Separable Computation of Information Measures.
• Wireless Communication. Learning-Based Receiver Design, MILCOM 2024.
• Feature Adaptation. Separable Design for Feature Regularization and Adaptation, Allerton 2024.
• Operator Learning. Learning-Based Operator SVD Solver, ICML 2024.
• Sequential Dependence. Feature Learning for Decomposing Sequential Dependence, Allerton 2023.
• Kernel Methods. Understanding Kernel Methods and Quantifying Kernel Quality, ISIT 2023.

Tutorials

The Geometric Information Learning blog illustrates the basic ideas, including PyTorch demos.

Selected Talk

Deep Learning From an Information Perspective

• Info
• Watch Here