Just a few years ago, whether in big companies or start-up companies, they are few engineers and scientists to study the depth applied to smart products and services. As the predecessor depth study of neural networks, machine learning is just to get rid of the impression that academics considered outdated tools. At that time, nor even the machine learning headlines regulars. It merely be seen as a forward-looking, and has a series of small-scale practical application of discipline. In practical applications include computer vision and natural language processing, including related areas usually requires a lot of knowledge: the practical application of these areas are regarded as independent of each other, and machine learning and only a small part.
However, only within the past few years, the depth of learning will make the whole world by surprise. It is very strong impetus to computer vision, natural language processing, automatic speech recognition, and strengthen the rapid development of the fields of learning and statistical modeling. With the continuous progress in these areas, we can now manufacture cars that drive themselves, even telephone automated response system, and beat the best human players in the Go software-based messaging, e-mail. These brought about by deep learning new tools are generating a broad impact: they changed the diagnosis filmmaking and the way and play an increasingly important role in everything from astrophysics to biology and other basic sciences.
At the same time, the depth of learning but also to bring its users a unique challenge: any single application brings together all the various disciplines of knowledge. In particular, the application need to understand the depth of learning:
Motivations and characteristics of the problem;
- The mathematical principles behind the model number of different types of neural network layers together through a specific manner;
- Fitting optimization model very complex deep in the raw data;
- Effective training model, avoid pitfalls and take advantage of numerical computation required hardware performance engineering skills;
- The selection of the appropriate variable (hyper-parameters) for the combined solutions of experience.
Similarly, we have several authors also facing unprecedented challenges: We need to blend aspects of knowledge in depth study of the limited space, so that readers can quickly understand and apply deep learning technology. Our book represents an attempt: We will teach the reader to the concept of background knowledge and code; we will explain critical thinking required to analyze the problem in the same place, the knowledge needed to solve mathematical problems, and implement solutions required engineering skills.
Why read this book:
- Each chapter of the book with the text, math, illustrations and code to introduce a wide range of knowledge points. It is a Jupyter notepad, you can run independently. It contains approximately 20 characters and code blocks can be read about 15 minutes. (See below)
- The source file is Markdown, do not save the output execution, and open source on Github. So easy for people to contribute and review the changes, and can easily continue to add new chapters.
- Any changes will trigger a continuous integration service to re-execute the implementation of Notepad to get the output, thus ensuring the correctness of the code. A notepad control execution time not to exceed ten minutes. This is quite a challenge to show training complex models.
- After correct execution directly available online in three formats: Jupyter notepad contains the implementation of output, you can directly browse the HTML, and suitable for printing PDF.
- It can be as easy as LaTeX index figures, tables, equations and documentation.
- Each chapter has a link that can be discussed.
With the famous Goodfellow, and compare Bengio Courville "deep learning" and Benefits
Goodfellow, Bengio Courville and "deep learning." The book reviews the many concepts and methods behind the depth of learning, is a very good teaching material. However, these resources are not described in conjunction with the concept of the actual code that can sometimes make the reader feel no idea how to implement them. In addition to these, providing business courses who despite making a large number of high-quality resources, but their paywall so many users daunting.
Because of this, deep learning users, especially beginners, often have to refer to a variety of different sources of information. For example, the control algorithm and related mathematical knowledge through textbooks or papers, read the online documentation to learn to use deep learning framework, and then look for algorithm interested in this framework, and to explore how to apply it to your own projects go with. If you are experienced in this process, you may feel pain: data from different sources is sometimes difficult to correspond with each other, even if the correspondence can also be required to spend a lot of energy. For example, we will need some papers in mathematical formulas and variables certain online implementation of the program variables to-one correspondence, and papers found in the code may not explain clearly the implementation details, and even install a different code for different operating operating environment.
For more than pain points exist, we are working to create a uniform resource to achieve the following objectives:
- Anyone can get for free on the Internet;
- Provide sufficient technical depth to help readers become actual depth learning applications scientists - both to understand the principles of mathematics, but also to achieve continuous improvement and methods;
- Contains the code can run, shows readers how to solve the problem in practice, not only directly correspond to mathematical formulas into actual code, and can modify the code, observations and timely access experience;
- And the entire community allows us to continue to fast iterative content to keep up with the rapid development of deep learning is still in the field;
- A Q & A forum contains information about technical details as a supplement, so that we may answering each other and exchange experiences.
Amazon chief scientist Li Mu and other masterpieces
Amazon book four authors are scientists, rather reputation in the field of artificial intelligence, Li Mu teacher is more like many fans from his mouth open class, pure Chinese teaching methods have been widely recognized by everyone.
Zhang Aston (Aston Zhang)
Amazon applied scientist, Dr. Urbana-Champaign University of Illinois computer science, statistics and computer science double master. His research focused on machine learning, and published in several top academic conferences over the papers. He served as NeurIPS, ICML, KDD, WWW, WSDM, SIGIR, AAAI and other academic conference program committee or reviewers and editorial board of the journal Frontiers in Big Data.
Li Mu (Mu Li)
Amazon Chief Scientist (Principal Scientist), Berkeley, visiting assistant professor at the University of California, Department of Computer Dr. Carnegie Mellon University. Distributed systems and machine learning algorithms to focus on him. He is one of the authors of the depth learning framework MXNet. He served as machine learning start-up companies Marianas Labs CTO and Baidu's director of R & D Institute of deep learning architect. In his theory, machine learning, fields of application and operating system and other top academic conferences (including FOCS, ICML, NeurIPS, AISTATS, CVPR, KDD, WSDM, OSDI) had published papers.
Zachary · C. Lipton (Zachary C. Lipton)
Amazon applied scientist, an assistant professor at Carnegie Mellon University, Ph.D., University of California, San Diego. He specializes in machine learning algorithms and their social impact, especially in depth study on the timing and sequence data decisions. Such work has a wide range of scenarios, including medical diagnosis, the dialogue system and product recommendations. He founded the blog "Approximately Correct" ( http://approximatelycorrect.com ).
Alexander · J. Mora (Alexander J. Smola)
Amazon vice president / eminent scientist, Dr. Computer Science Technical University of Berlin, Germany. He studied at the Australian National University, University of California, Berkeley, and Carnegie Mellon University. He has published more than 200 papers and five books, his papers and books have been cited more than 100,000 times. His research interests include deep learning, Bayesian nonparametric, nuclear methods, statistical modeling and scalable algorithms.
Content and structure
This book can be roughly divided into three parts.
The first part (Chapter 1 to Chapter 3) covers the preparatory work and the basics. Chapter 1 introduces the background depth learning. Chapter 2 provides hands-on science learning depth prior knowledge required, for example, how to obtain and run the code in this book. Chapter 3 includes deep learning the most basic concepts and techniques, such as MLP model and regularization. If the reader time is limited, and just want to know the depth of learning the most basic concepts and techniques, then just read the first part.
The second part (Chapter 4 to Chapter 6) the depth of focus of modern learning technologies. Chapter 4 describes an important part of the calculation of the various depth study and lay the foundation for more complex models follow. Chapter 5 explains the depth of study in recent years make a big success of convolution neural networks in the field of computer vision. Chapter 6 describes commonly used in recent years, the neural network processing loop sequence data. Read the second part helps to grasp the depth of modern learning technologies.
The third part (Chapter 7 - Chapter 10) discusses computational performance applications. Chapter 7 evaluate various optimization algorithms used to train the depth learning model. Chapter 8, several important factors affect test depth study computing performance. Deep learning important applications in computer vision and natural language processing in Chapter 9 and Chapter 10 are listed. This part of the reader can choose according to interest in reading.
Figure 0-1 depicts the structure of the book, which chapter by A to point B of Chapter A chapter of the arrow indicates that knowledge helps to understand the content of B Chapter.
Figure 0-1 Structure of the book
"Hands-on learning deep learning (Paperback)"
Author: Zhang Aston (Aston Zhang) Li Mu (Mu Li) [US] Zachary C. Lipton (Zachary C. Lipton) [Germany] Alexander J. Mora (Alexander J. Smola)
Editor's Choice
- Amazon scientist Li Mu and other heavy work;
- The new mode of interactive hands-on learning environment real depth of learning, the perfect combination of theory and actual combat
- Jiawei Han / Bernhard Schölkopf / Zhou Zhihua / Zhang Tong / Yu Kai / paint away / Shen Qiang jointly recommended
- University of California at Berkeley and other 15 well-known universities around the world for teaching
This book aims to deliver learning about the depth of interactive learning experience to the reader. The book not only describes the principles of deep learning algorithms, but also demonstrate their implementation and operation. Traditional books different, each section of the book is a Jupyter can download and run the Notepad, it text, formulas, images, code and operating results combined together. In addition, readers can also access the contents of the book and participate in discussions.