دانلود کتاب یادگیری ماشین قابل اعتماد تحت داده‌های ناقص

Machine learning has become increasingly prevalent in decision-critical applications. However, real-world deployment scenarios often deviate from idealized assumptions where data are perfectly labeled, identically distributed, and free from malicious interference. Therefore, ensuring the trustworthiness of machine learning has emerged as one of the most pressing challenges in machine learning research. In practice, machine learning systems frequently encounter imperfect data in three main forms: noisy labels due to annotation errors, adversarial examples crafted to mislead predictions, and out-of-distribution samples that differ from training data. These challenges represent critical vulnerabilities that can lead to catastrophic failures, such as misdiagnosis in medical imaging, fraudulent transactions bypassing financial safety systems, and safety hazards in autonomous vehicles. This book provides a comprehensive exploration of trustworthy machine learning in imperfect data scenarios. Through six structured chapters, we present a unified framework encompassing the major categories of data imperfections and their mitigation strategies. Chapter 1 establishes the fundamental concepts of trustworthy machine learning and provides a comprehensive taxonomy of imperfect data scenarios. We introduce the core principles of robustness, safety, and reliability that define trustworthiness, while categorizing the three primary types of imperfect data: noisy labels, adversarial examples, and out-of-distribution samples. Chapter 2 focuses on trustworthy machine learning with noisy labels, exploring statistical learning approaches, deep learning methodologies, applications to overparameterized models, and the natural transition from noisy label learning to weakly supervised learning paradigms. We examine how different noise assumptions— symmetric, class-dependent, and instance-dependent noise—lead to distinct theoretical properties and practical solutions, providing both theoretical guarantees and empirical validation. Chapter 3 addresses adversarial examples from three complementary perspectives: statistical learning methods that model the transition relationships between adversarial and natural labels, deep learning approaches that enhance robustness through adversarial training, and diffusion-based methods that purify or detect adversarial perturbations in the preprocessing stage. This multifaceted approach provides comprehensive coverage of defense strategies throughout the entire machine learning pipeline. Chapter 4 examines out-of-distribution data challenges through two perspectives: detection and generalization. In out-of-distribution detection, we explore methods for identifying when test samples fall outside the training label space, enabling models to abstain from making unreliable predictions. In out-of-distribution generalization, we investigate techniques to improve generalization across different domains and distributions, focusing on invariant risk minimization and contrastivebased approaches. Chapter 5 extends our exploration to advanced application domains that represent the cutting edge of trustworthy machine learning research: federated learning, graph learning, and foundation models. These domains present unique challenges that amplify the importance of trustworthiness considerations. Federated learning introduces privacy and safety concerns in distributed settings, graph learning requires handling complex relational data structures with novel vulnerability patterns, and foundation models demand trustworthiness at unprecedented scale. Chapter 6 synthesizes the key insights from previous chapters and provides a forward-looking perspective on the future directions of trustworthy machine learning research. We discuss emerging challenges, identify promising research opportunities, and offer practical guidance for implementing trustworthy machine learning systems in real-world applications. We hope this book provides an introductory background to the concept and coverage of trustworthy machine learning and summarizes some of the cutting-edge progress in trustworthy machine learning research. This can prepare the readers to embark on a fast track to keep up with the current dilemmas and challenges in this research area, conducting new research upon existing methods, and adapting relevant heuristics to scenarios that require trustworthy machine learning systems. Finally, we would like to thank the members of the Sydney AI Centre, TML Lab, and TMLR Group who have contributed to this book. They are: Yingbin Bai, Zhuo Huang, Muyang Li, Xiuchuan Li, Runqi Lin, Yuhao Wu, Yu Yao, Chaojian Yu, Suqin Yuan, and Dawei Zhou from the Sydney AI Centre and TML Lab; and Chentao Cao, Yongqiang Chen, Haoang Chi, Rong Dai, Zhen Fang, Xiao Feng, Xue Jiang, Xuan Li, Chenxi Liu, Feng Liu, Xiangyu Lu, Zhenheng Tang, Qizhou Wang, Binghui Xie, Zhiqin Yang, Jiangchao Yao, Shuhai Zhang, Yonggang Zhang, Zhun Zhong, Zhanke Zhou, and Jianing Zhu from the TMLR Group (listed alphabetically within each group).