Deep non-IID learning
Zhilin Zhao, Longbing Cao. IJCAI 2023. Slides

Deep neural networks demonstrate great success in many real-world applications with the basic assumption that samples are independent and identically distributed (IID). However, data goes beyond the IID assumption owing to various complexities, i.e., non-independent and -identically distributed (non-IID), and the heterogeneity and interaction form the comprehensive non-IIDness. Accordingly, non-IID data is common in many deep learning applications, including out-of-distribution detection, distribution discrepancy estimation, domain generalization, representation learning, federated learning, and multiple time-series analysis. Due to the non-IID nature of complex data, it is essential to explore the explicit and implicit interactions and couplings embedded in heterogeneous and nonstationary data for deep learning to adapt to real-world applications.

Deep Non-IID Learning explores the non-IIDness in complex data for addressing issues in different learning paradigms. This tutorial is dedicated to illustrating the frameworks of deep non-IID learning, introducing advanced methods for real-world applications from a deep non-IID learning perspective, and discussing the challenges and opportunities. The potential audience will be machine learning researchers and industry practitioners who are interested in exploring complex data characteristics and addressing the issues in real-world applications by deep learning. The tutorial aims to enable both academic and practical audiences with a comprehensive understanding of how the specific research issues of complex real-world applications can be addressed by exploring the non-IIDness in deep neural networks.

Access the relevant information on non-IID learning at the non-IID learning webpage.