دانلود کتاب مقدمه‌ای بر مدل‌های فونداسیون

Back in June 2022, we (Pin-Yu Chen, Sijia Liu, and Sayak Paul) sensed that foundational models and generative AI were about to bring unprecedented and revolutionary changes to our technology and society, and we decided to propose a tutorial at NeurIPS to educate researchers and practitioners about this emerging AI research and technology and its implications for robustness and safety. Two months later, our tutorial, “Foundational Robustness of Foundation Models,” was accepted and to be presented on December 5, 2022. At that time, the world was still recovering from the Covid crisis, and NeurIPS 2022 was the first NeurIPS conference to allow physical attendance after Covid. To accommodate the hybrid conference mode, we were asked to record our tutorial by October 2022. In our materials, we correctly predicted that the proliferation of foundation models would bring increased risks and challenges in robustness, security, and safety, later known as part of the broad topic of AI safety, governance, and risk assessment. About a week before the official release of our tutorial and the start date of NeurIPS 2022, ChatGPT was launched by OpenAI on November 30, 2022. At the conference, we heard only a few whispers and genuine inquiries about “Have you heard of ChatGPT?” Who would have thought that in August 2024, the number of ChatGPT users worldwide would be more than 180 million? With such a coincidence in mind, we felt compelled to expand our tutorial into a book, summarizing the basics of foundation models and generative AI, and emphasizing the robustness and safety of these frontier models and AI technology. This book consists of three parts. Part I (Chaps.1 to 4) provides the fundamentals of foundation models, Part II (Chaps.5 to 11) includes advanced topics in founda-tion modes, and Part III (Chaps.12 to 18) presents safety and trust in foundation models. The mathematical notations will be defined and explained in each chapter. We provide an overview of each part as follows: • In Part I, Chap.1 provides an overview of foundation models and generative AI. Chapter 2 presents the technical background of neural networks. Chapter 3 delves into the learning and generalization of transformers. Chapter 4 formalizes in-context learning with transformers. • In Part II, Chap.5 introduces automated visual prompting techniques. Chapter 6 introduces prompting LLMs with privacy. Chapter 7 elucidates memory-efficient fine-tuning methods. Chapter 8 shows how LLMs can be reprogrammed for time-series machine learning tasks. Chapter 9 shows how LLMs can be reused for speech tasks. Chapter 10 introduces how synthetic datasets can be used to benchmark foundation models. Chapter 11 elucidates machine unlearning for foundation models. • In Part III, Chap.12 provides a comprehensive evaluation of the trustworthiness of LLMs. Chapter 13 introduces jailbreak attacks and defenses for LLMs. Chapter 14 presents safety risks when fine-tuning LLMs. Chapter 15 introduces watermarking techniques for LLMs. Chapter 16 presents robust detection of AI-generated text. Chapter 17 elucidates backdoor risks in diffusion models. Chapter 18 presents red-teaming methods for diffusion models. The authors would especially like to thank Paul Drougas at Springer for his encouragement and patience in allowing us to publish a book in such a rapidly changing and growing field. Some parts of the book are based on publications co-authored by the authors of this book. We would like to express our gratitude to our research collaborators. Finally, we would like to thank our families, colleagues, and friends, for their support.