Learning from Data with the Maximum Correntropy Criterion

Robust machine learning and signal processing beyond mean square error for real-world non-Gaussian data and outliers.

This essential resource for researchers and graduate students in signal processing and AI introduces correntropy-based learning methods that outperform traditional techniques in challenging environments involving noisy real-world data. Topics covered include adaptive filtering, neural networks, and machine learning.Traditional signal processing and machine learning methods rely on mean square error, which becomes brittle when data contains heavy-tailed noise, impulsive disturbances, and outliers—conditions frequently encountered in real-world applications. This comprehensive guidebook introduces correntropy-based methods that demonstrate superior robustness across diverse engineering domains, progressing from foundational concepts to applications. Authored by pioneers in information theoretic learning, the book systematically covers correntropy fundamentals, adaptive filtering techniques, neural network training, feature learning, and applications including point set registration, matrix completion, and federated learning. Each chapter balances rigorous theory with practical algorithms and performance comparisons against conventional methods. With implementation guidelines and a unified framework connecting different robust learning criteria, this book addresses the critical gap between Gaussian-assumption theory and non-Gaussian reality, providing researchers and graduate students with applicable solutions for challenging real-world problems.