Translator's Preface ix
To Korean readers x
Preface xi

CHAPTER 1 Generative AI 1

What is Generative AI? Part 1
Creation according to instructions and conditions 2
Ability to create data that was previously difficult to create 3
From Rule-Based to Machine Learning 5
Generation tasks are a particularly difficult machine learning problem 7.
Generating data is like finding an island in a vast ocean.
Vast and Strange High-Dimensional Space 11
Generation has more than one correct output 13
Manifold Hypothesis: Data in Low Dimensions 15
Symmetry: Data that is invariant to transformations 18
Composition: Data made up of a combination of multiple parts 20
[COLUMN] Are data's characteristics provided by humans, or are they self-learning? 21
Summary 22

CHAPTER 2: A History of Generative AI 23

Mechanisms of Memory 23
From the Easing model to the Hopfield network 24
Energy-based model 28
Energy-Based Model 29 for Naturally Realizing Associative Memory
The Relationship Between Energy and Probability: Boltzmann Distribution 31
Principles of the Langevin Monte Carlo Method 32
33 Fatal Problems with Energy-Based Models
[COLUMN] The Real World is a Giant Simulator 34
Distribution function 35 that governs information throughout space
Data generated from hidden information 37
Awareness is required for creation 38
Variational Autoencoder (VAE) 40
Problem 42 of the Latent Variable Model
[COLUMN] Generative Adversarial Networks (GANs) 43
[COLUMN] Autoregressive Model 43
[COLUMN] 2024 Nobel Prize 44
Summary 45

CHAPTER 3 Creating Using Flow 47

Floran 47
Continuity Equation: Matter Does Not Suddenly Disappear or Warp 49
Complex probability distributions created using flow 51
Flow-based model 53 that does not require a distribution function
Normalization Flow and Continuous Normalization Flow 55
Learning to maximize the likelihood obtained along the flow 55
Generate data according to flow 57
Flow 58: Decomposing complex generation problems into simpler sub-generation problems
Flow Modeling 60
Flow result calculation 62
Normalization Flow Challenge 64
Summary 65

CHAPTER 4 Diffusion Models and Flow Matching 67

Discovery of the Diffusion Model 67
General diffusion phenomenon 68
[COLUMN] Brownian Motion 69
The diffusion model is 70
Flow created by the diffusion process = Score 72
The Relationship Between Score and Energy 73
Score 74, which changes with time
Denoising Score Matching 76
Simulation-free learning is only available to some 78
Summary of Learning and Generation by Diffusion Models 79
Characteristics of flow generated by diffusion models 79
Relationship between diffusion models and latent variable models 80
Automatically learning phylogenetic trees for data generation 81
The diffusion model is an energy-based model 82
The diffusion model is a generative model using flow 82
Flow Matching: Complex Flow 83 created by collecting flows
Optimal Transportation 83
Generation 85 using optimal transport
Finding the optimal transport directly is too computationally intensive 85
Learning Flow Matching 86
The evolution of flow matching 88
Conditional generation is realized with conditional flow 88
Latent Diffusion Model: Transforming Original Data into Latent Space to Improve Quality 90
Summary 91

CHAPTER 5: THE FUTURE PROSPECTS OF FLOW-BASED TECHNOLOGY 93

Solving the Mystery of Generalization 93
Generation 95 with symmetry in mind
Attention Mechanism and Flow 96
Numerical Optimization by Flow 96
Generating discrete data such as language 97
99 Contact with the brain's computational mechanisms
The Future of Creation by Flow 99

APPENDIX A Machine Learning Keywords 101

Probability and Generative Models 101
Maximum likelihood method 102
Machine Learning 103
Machine Learning Mechanisms 104
Parameter tuning = learning 105
Neural Network 106
Generalization 106: Obtaining rules applicable to infinite data from finite training data

APPENDIX B REFERENCES 109

Chapter 2, 110
Chapter 3, 112
Chapter 4, 112
Chapter 5, 114

Search 117

A book explaining generative AI using only text and pictures.

Flow-based generative techniques, especially diffusion models, have emerged in many fields, including image, audio, and video generation.
Daisuke Okanohara, Japan's top AI expert, wrote "The Mathematics of Diffusion Models", which clearly explains the diffusion model mathematically, and this time he wrote "The Structure of Generative AI", which explains the entirety of generative AI using only text and pictures, without any mathematical formulas.


It is written in an easy-to-read format that even non-experts can understand, covering everything from the history of generative AI to flow, diffusion models and flow matching, optimal transportation, and future prospects.
Instead of avoiding formulas, I prefer to use everyday metaphors from the world we live in. For example, I begin my explanation of the diffusion model as follows:


“Let’s say you write with ink on the surface of water.
The letters written with this ink will gradually dissolve over time, and eventually the ink will mix evenly throughout the water.
(…) If we could reproduce this ink diffusion process in the opposite direction, we could revert from a state where the ink was evenly mixed in water to a state where the letters were written in ink again.
In other words, the idea is that by reversing the process of adding noise to an object of order, gradually destroying it and turning it into complete disorder, we can create order from disorder—that is, create it.

Daisuke Okanohara is famous as the co-founder of Preferred Networks, Japan's largest AI unicorn, but he is also a veteran author who has written over a dozen professional books and is praised for "providing rich context surrounding technology."
Let's properly understand the structure of generative AI, which is at the center of today's IT, with his kind and accurate explanation.