Public account [Machine Learning and AI Generated Creation] Backstage reply : 168 . You can participate in the free delivery of [Deep Generative Model] book activity. The deadline for the activity is 10 pm on September 17, 2023 .
You can also purchase it yourself by clicking on the following JD.com exclusive link:
What I cannot create, I do not understand.
——Richard Feynman
This sentence from Feynman, the Nobel Prize winner in physics, has been ringing in the hearts of artificial intelligence researchers. OpenAI's research has always been adhering to a core desire, which is to "develop algorithms and technologies to let computers understand us." world."
There are currently two ways for computers to understand our world, namely supervised learning and unsupervised learning .
Among them, although supervised learning has made great technical progress, its biggest shortcoming is that the data annotation process is often time-consuming and expensive, may even be unethical, or even impossible. This is why we are increasingly aware of the importance of unsupervised learning (or self-supervised learning) methods.
Unsupervised learning is the most suitable model for human learning knowledge. Just like when a child grows up, no one will always tell them what the information they see and hear is. Most of the information they are exposed to about the world is Unlabeled, you need to learn the laws of the world in an unsupervised manner, that is, find patterns and structures in the data yourself.
The deep generative model is one of the most effective methods of unsupervised learning and is also one of the most promising methods to achieve general artificial intelligence according to OpenAI.
The book "Deep Generative Models" covers various types of deep generative models, including autoregressive models, flow models, latent variable models, energy-based models, etc. It mainly talks about how to combine probabilistic modeling and deep learning to build A powerful AI system that can quantify uncertainty in the surrounding environment.
This kind of AI system can understand the surrounding world from a generative perspective.
These models mentioned in the book form the technical cornerstone behind deep generative models such as large language models represented by ChatGPT and diffusion models represented by Stable Diffusion.
As the translator said, translating this book brought him out of the heat, anxiety and confusion brought about by the emergence of ChatGPT, and instead made him think more calmly. Therefore, for new technologies, only by understanding the principles behind them can people feel comfortable with them.
Although we may not be able to master all new technologies such as large language models, most of the basic content behind them is reflected in this book. As long as you read this book carefully, derive formulas, run codes, and learn the methods of various generative models Development and its application will allow us to understand and face new technologies more calmly.
The Transformer architecture behind ChatGPT and its related models from GPT-1 and GPT-2 to BERT and GPT-3 are all developed based on the autoencoder described in Chapter 4 of this book and the autoregressive model described in Chapter 2 Come. ChatGPT and other related large language models mainly focus on data collection, data scale and quality, and model parameter expansion, but their generative model infrastructure is still based on that described in this book, but the language model is optimized.
In addition to the above-mentioned generative models of text-to-image or text-to-text, more multi-modal applications are developing in full swing, such as from text to 3D model (Dreamfusion, Magic3D), video (Phenaki, Soundify), voice (AudioLM, Whisper, Jukebox), code (CodeX, Alphacode), and models for generating text from images (Flamingo, VisualGPT), etc. The technical cornerstones of these applications can be found in this book.
The special thing about this book is that just like children learn the laws of physics by building blocks, readers of this book can learn deep generative models through code.
The author of this book has published a large number of papers in this field and has a deep understanding of this field.
This book can be used as a textbook for courses on probabilistic generative models.
