"The Age of Intelligence: Big Data and the Intelligent Revolution Redefines the Future" predicts that traditional industries will use intelligent technology to achieve upgrading, and will also change the original business model. Indeed, in the past two years, the innovation of artificial intelligence in the application side has mushroomed, and AI applications such as unmanned driving and intelligent recommendation have become a reality.
But at the same time, we also found that most of the landing scenarios of AI originate from capital-intensive industries, which is not in line with the background of AI universal benefit and service industry intelligence.
In fact, what is hindering the development of the AI industry is neither algorithms nor computing power, but how to better transform AI capabilities into engineering means. In addition to the well-known core application scenarios such as automatic driving, there are also a large number of engineering fields, complex scenarios in thousands of industries, and various terminal equipment, which require the assistance of AI technology.
Recently, I read a book in the Polar Arts Community (aijishu.com), "AI Embedded Systems: Algorithm Optimization and Implementation", which can help students or developers of embedded technology-related majors improve their skills and accelerate the large-scale commercial landing of embedded AI in the industry.
Embedded AI provides theoretical support for AI commercialization
In the process of AI commercialization, to achieve the landing of technology, it is not only necessary to have a superior algorithm model and reliable hardware support, but also to organically combine AI technology and hardware environment and then apply it to specific practical scenarios. If the future is an intelligent world, then every terminal should be intelligent, which must rely on embedded AI technology.
Embedded refers to a dedicated computer system that can be built into a device or device. In general, devices with digital interfaces have embedded systems, such as mobile phones, car computers, smart watches, and so on. Embedded AI, on the other hand, is a technology that allows AI algorithms to run on end devices
In general, embedded AI has a natural advantage for high real-time processing scenarios, such as drones, smart cars, industrial machinery, electrical appliances, security inspection and other scenarios have been more and more applications.
Many people may ask, after the maturity of 5G networks, AI capabilities can be uploaded to the cloud for processing, why embedded AI? In fact, the scene in which the terminal is located is more complex, such as intelligent car driving will encounter unpredictable road conditions, it is impossible to ensure that the data can be synchronized in time, to make the car make intelligent driving or even unmanned timely response, it needs embedded AI, so that it can make judgments in time.
From the developer's point of view, learning AI embedded systems is also difficult. For example, how to train a neural network to recognize patterns, how to optimize the network to meet the limited size of embedded devices, so as to achieve size and power goals, and so on. It's a systematic job.
Therefore, "AI Embedded Systems: Algorithm Optimization and Implementation" starts from the underlying computing details of deep neural networks and provides a comprehensive and in-depth interpretation of embedded AI technology. It can be used as a reference textbook for senior undergraduate students in computer, electronic information, automatic control and other related majors for engineers and technicians in the embedded field, and the underlying algorithm software developers of machine learning algorithms.
Combining theory and Practice: Choices under Different Paths
Many series of books in the computer industry will have the old problem of too much emphasis on theoretical knowledge and lack of practical support.
"AI Embedded System: Algorithm Optimization and Implementation" is a set of machine learning algorithm optimization theory and application practice as a whole, not only for the embedded application of artificial intelligence, covering the machine learning algorithm optimization theory and technology based on multiple dimensions such as operation volume, memory, power consumption and so on. From the underlying unit of the algorithm to the joint optimization of machine learning algorithm training-inference and automatic optimization deployment algorithm, it is also verified in combination with the general embedded processor.
From the choice of processor system architecture and the choice of machine learning algorithms, the practical value of this book can also be seen in both directions.
Now many machine learning algorithms are based on massive storage and computing, which have higher requirements for power consumption, volume, computing power and storage capacity, and embedded systems in different fields have multiple ways, in specific scenarios, how to choose, which increases the difficulty of developer learning to a certain extent.
For example, the general multiprocessor architecture based on the GPU has strong versatility and high computing power, but the price is high hardware cost and large power consumption; the customized solution based on the dedicated computing acceleration engine has high computational efficiency, controllable power consumption, but the computing structure is relatively fixed and the flexibility is poor; and the scheme based on the existing processor for in-depth optimization of the algorithm is low cost, does not require special or customized hardware, and has strong versatility, but the cost is that it needs to be manually optimized one by one for each case, and the development difficulty is high.
Existence is reasonable, and different paths have their existence value in different scenarios.
AI Embedded Systems: Algorithm Optimization and Implementation focuses on the implementation of machine learning algorithms based on existing embedded processor systems and architectures through different levels of optimization. Although the methods described take the general-purpose embedded processor as an example, they can also be applied to GPU or hardware acceleration engine architectures, such as fast convolution algorithms based on transformation domains, which are applied in the underlying computing library of the GPU, and constant multiplication operations based on addition and subtraction graphs can be easily implemented in chip hardware design.
Diversity lies not only in processor systems and architectures, but also in machine learning and in several different areas of algorithm selection.
For example, the algorithm theory model based on the Department of Statistics is well designed, has good generalization, and the amount of operation is relatively low, which is easy to implement in an embedded system with limited storage amount of operation, but it relies on strong probability assumptions, and the deviation on the model limits its performance in practical applications.
In contrast, neural network algorithms, despite their superior performance, have a huge amount of parameter storage that limits their implementation on embedded systems. In recent years, research has found that there are a large number of redundant operations and parameters in many neural network architectures, and the simplification and architectural clipping of operations can greatly reduce the demand for computing performance and memory of neural networks, making it possible to implement it on embedded systems.
Therefore, the content discussed in AI Embedded Systems: Algorithm Optimization and Implementation also covers the implementation of statistically and neural network-based machine learning algorithms on embedded systems, and introduces specific optimization methods through algorithms and routines.
"Careful thinking" helps readers absorb knowledge points faster
Why is "AI Embedded Systems: Algorithm Optimization and Implementation" available to meet the needs of both developers and professional students?
First of all, in terms of professionalism, the two authors of this book have strong professional backgrounds, such as Ying Nindong, who is an associate professor at the School of Electronic Information and Electrical Engineering of Shanghai Jiao Tong University and a master tutor. He has long been engaged in the teaching and scientific research of embedded systems and digital circuits; Liu Peilin is a professor and doctoral supervisor of the School of Electronic Information and Electrical Engineering of Shanghai Jiao Tong University. They have both a professional perspective and know how to teach and solve puzzles.
Secondly, from the detailed content, the content of this book covers the general embedded optimization technology, including the optimization of SMD instruction set, memory access mode optimization, parameter quantization, etc., and on this basis, the optimization of signal processing level, the optimization of Al inference algorithm and the optimization theory and method of Al algorithm training-inference joint based on neural network are introduced. In addition, this book also introduces the basic knowledge of automatic optimization and deployment tool development of machine learning algorithms in a general embedded environment through a number of examples of automatic search for optimization parameters and generation of C code, and explains the implementation method of Al algorithm in general embedded systems through application examples and a large number of codes, and strives to let readers master the knowledge and skills of efficient Al embedded system development through practice on the basis of understanding algorithms.
Third, "AI Embedded Systems: Algorithm Optimization and Implementation" itself also takes into account the acceptance of different audiences, so the content of machine learning algorithms in embedded systems not only gives manual optimization steps, but also introduces software-aided optimization, so that readers can understand according to actual needs. In particular, the underlying algorithm code in the book is presented in python language form. This is to take into account Python's language expression ability and procedural simplicity, and these "careful thoughts" can help readers better absorb and learn the knowledge points of this book.
With the popularization of artificial intelligence technology in various industries, the combination of AI technology and embedded systems to build AI embedded systems has also laid the foundation for the large-scale landing of AI commercialization, which is also the original intention of the book "AI Embedded Systems: Algorithm Optimization and Implementation".