A0740
Title: Information theoretic learning meets deep neural networks
Authors: Jun Fan - Hong Kong Baptist University (Hong Kong) [presenting]
Abstract: Information theoretic learning, a machine learning approach that incorporates ideas from information theory, offers a family of supervised learning algorithms based on the principle of minimum error entropy (MEE). These algorithms provide an alternative to traditional least squares methods, particularly effective when dealing with heavy-tailed noises or outliers. The integration of information-theoretic learning with deep learning has garnered significant attention in addressing the evolving challenges of modern machine learning. The theoretical exploration of MEE algorithms generated by deep neural networks is delved into in the context of regression tasks. The focus is on establishing fast learning rates for these algorithms when the noise satisfies weak moment conditions.
