ChatGPT : principles and architecture /

"ChatGPT: Principles and Architecture bridges the knowledge gap between theoretical AI concepts and their practical applications. It equips industry professionals and researchers with a deeper understanding of large language models, enabling them to effectively leverage these technologies in th...

Full description

Bibliographic Details
Main Author:	Cheng, Ge (Author)
Corporate Author:	ScienceDirect (Online service)
Format:	eBook
Language:	English
Published:	Amsterdam, Netherlands : Elsevier Inc., [2025]
Edition:	First edition.
Subjects:	ChatGPT. Computational linguistics. Artificial intelligence. Linguistique informatique. Intelligence artificielle. computational linguistics. artificial intelligence. Electronic books.
Online Access:	Connect to the full text of this electronic book

Table of Contents:

Front Cover
ChatGPT
Copyright Page
Contents
Preface
Main Content of the Book
Target Audience for This Book
Contact the Author
Acknowledgments
1 A new milestone in artificial intelligence-ChatGPT
1.1 The development history of ChatGPT
1.2 The capability level of ChatGPT
1.3 The technical evolution of large language models
1.3.1 Symbolism versus connectionism
1.3.2 Transformer
1.3.3 Unsupervised pretraining
1.3.4 Supervised fine-tuning
1.3.5 Human feedback reinforcement learning
1.4 The technology stack of large language model
1.5 The impact of large language models
1.6 The challenges of training or deploying large models
1.6.1 Computational power
1.6.2 Data
1.6.3 Engineering
1.7 The limitations of large language models
1.8 Summary
2 In-depth understanding of the transformer model
2.1 Introduction to the transformer model
2.2 Self-attention mechanism
2.2.1 The calculation process of self-attention
2.2.2 The essence of the self-attention mechanism
2.2.3 The advantages and limitations of the self-attention mechanism
2.3 Multihead attention mechanism
2.3.1 Implementation of multihead attention
2.3.2 The role of multihead attention
2.3.3 Optimization of multihead attention
2.4 Feedforward neural network
2.5 Residual connection
2.6 Layer normalization
2.7 Position encoding
2.7.1 Design and implementation of positional encoding
2.7.2 Variants of positional encoding
2.7.3 The advantages and limitations of positional encoding
2.8 Training and optimization
2.8.1 Loss function
2.8.2 Optimizer
2.8.3 Learning rate adjustment strategy
2.8.4 Regularization
2.8.5 Other training and optimization techniques
2.9 Summary
3 Generative pretraining
3.1 Introduction to generative pretraining.
3.2 Generative pretraining model
3.3 The generative pretraining process
3.3.1 The objectives of generative pretraining
3.3.2 The error backpropagation process in generative pretraining
3.4 Supervised fine-tuning
3.4.1 The principles of supervised fine-tuning
3.4.2 Supervised fine-tuning for specific tasks
3.4.3 Fine-tuning steps
3.5 Summary
4 Unsupervised multitask and zero-shot learning
4.1 Encoder and decoder
4.2 GPT-2
4.2.1 Layer normalization
4.2.2 Orthogonal initialization
4.2.3 Reversible tokenization
4.2.4 Learnable relative positional encoding
4.3 Unsupervised multitask learning
4.4 The relationship between multitask and zero-shot learning
4.5 The autoregressive generation process of GPT-2
4.5.1 Subword unit embeddings
4.5.2 Autoregressive process
4.6 Summary
5 Sparse attention and content-based learning
5.1 GPT-3
5.2 The sparse transformer
5.2.1 Characteristics of the sparse transformer
5.2.1.1 Sparse attention patterns
5.2.1.2 Alternating dense and sparse attention patterns
5.2.1.3 Learnable relative positional encodings
5.2.2 Local banded attention
5.2.3 Cross-layer sparse connections
5.3 Meta-learning and in-context learning
5.3.1 Meta-learning
5.3.2 In-context learning
5.4 Bayesian inference of concept distributions
5.4.1 Implicit fine-tuning
5.4.2 Bayesian inference
5.5 Thought chains
5.6 Summary
6 Pretraining strategies for large language models
6.1 Pre-training datasets
6.2 Processing of pretraining data
6.3 Distributed training patterns
6.3.1 Data parallelism
6.3.2 Model parallelism
6.4 Technical approaches to distributed training
6.4.1 Pathways
6.4.2 Megatron-LM
6.4.3 ZeRO
6.5 Examples of training strategies
6.5.1 Training framework
6.5.2 Parameter stability.
6.5.3 Optimizing training settings
6.5.4 BF16
6.5.5 Other factors
6.6 Summary
7 Proximal policy optimization
7.1 Traditional policy gradient methods
7.1.1 The principles of policy gradient methods
7.1.2 Importance sampling
7.1.3 Advantage function
7.2 Actor-Critic
7.2.1 Algorithm steps
7.2.2 Value function and policy update
7.2.3 Issues and challenges
7.3 Trust region policy optimization
7.3.1 Optimization objectives
7.3.2 Limitations
7.4 Principles of the proximal policy optimization algorithm
7.5 Summary
8 Human feedback reinforcement learning
8.1 Reinforcement learning in ChatGPT
8.2 InstructGPT training dataset
8.2.1 Sources of fine-tuning datasets
8.2.2 Annotation standards
8.2.3 Data analysis
8.3 Training stages of human feedback reinforcement learning
8.3.1 Supervised fine-tuning
8.3.2 Reward modeling
8.3.3 Reinforcement learning
8.4 Reward modeling algorithms
8.4.1 Reward scores
8.4.2 Loss function
8.5 PPO in InstructGPT
8.6 Multiturn dialogue capability
8.7 The necessity of human feedback reinforcement learning
8.8 Summary
9 Low-resource domain transfer of large language models
9.1 Self-instruct
9.1.1 Instruction generation
9.1.2 Task classification identification
9.1.3 Instance generation
9.1.4 Filtering
9.2 Constitutional artificial intelligence
9.3 Low-rank adaptation
9.3.1 Model training and deployment
9.3.2 Choice of rank
9.4 Quantization
9.5 SparseGPT
9.6 Case studies
9.6.1 Base model
9.6.2 Instruction-following model
9.6.3 Medical field
9.6.4 Judicial field
9.7 Summary
10 Middleware
10.1 LangChain
10.2 AutoGPT
10.3 Competitors in middleware frameworks
10.4 Summary
11 The future path of large language models
11.1 The path to strong artificial intelligence.
11.2 Data resource depletion
11.3 Limitations of autoregressive models
11.4 Embodied intelligence
11.4.1 Challenges of embodied intelligence
11.4.2 PaLM-E
11.4.3 ChatGPT for robotics
11.5 Summary
Index
Back Cover.