In the book “The Intelligent Era: Big Data and the Intelligent Revolution Redefining the Future”, it is predicted that traditional industries will adopt intelligent technologies for upgrades, which will change existing business models. Indeed, in the past two years, innovations in artificial intelligence at the application end have sprung up like mushrooms after rain, with applications such as autonomous driving and intelligent recommendations becoming a reality.However, we also find that most AI implementation scenarios originate from capital-intensive industries, which does not align with the background of AI’s inclusiveness and service to industrial intelligence.In fact, what hinders the development of the AI industry is neither the algorithms nor the computing power, but how to better convert AI capabilities into engineering methods. In addition to well-known core application scenarios like autonomous driving, there are many engineering fields with complex scenarios in various industries that require the assistance of AI technology.Recently, I read a book titled “AI Embedded Systems: Algorithm Optimization and Implementation” in the Extreme Technology Community (aijishu.com), which can help students or developers related to embedded technology improve their skills and accelerate the large-scale commercial implementation of embedded AI in the industry.
Embedded AI as Theoretical Support for AI CommercializationIn the process of AI commercialization, to achieve the implementation of technology, not only superior algorithm models and reliable hardware support are needed, but also an organic combination of AI technology and hardware environments, which should then be applied 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 devices or apparatus. Generally speaking, devices with digital interfaces have embedded systems, such as mobile phones, vehicle computers, smartwatches, etc. Embedded AI is a technology that allows AI algorithms to run on terminal devices.Typically, embedded AI has inherent advantages in high real-time processing scenarios, such as drones, smart cars, industrial machinery, electrical appliances, and security inspections, where it is increasingly applied.Many may ask, after the maturity of 5G networks, why use embedded AI when AI capabilities can be uploaded to the cloud for processing? In fact, the scenarios where terminals operate are quite complex. For example, when a smart car is driving, it may encounter unpredictable road conditions, which makes it impossible to ensure timely data synchronization. To enable the car to make intelligent driving or even autonomous driving responses promptly, embedded AI is needed to make timely judgments.
From a developer’s perspective, learning about AI embedded systems also comes with many challenges. For instance, how to train neural networks to recognize patterns and how to optimize these networks to meet the size constraints of embedded devices to achieve size and power targets. This is a systematic task.Therefore, “AI Embedded Systems: Algorithm Optimization and Implementation” starts from the underlying computational details of deep neural networks and provides a comprehensive and in-depth interpretation of embedded AI technology. It addresses the concerns of engineering technicians in the embedded field and developers of underlying algorithms for machine learning, and can also serve as a reference textbook for senior undergraduate students in related majors such as computer science, electronic information, and automatic control.Combining Theory and Practice: Choices Under Different PathsMany computer industry series books tend to overemphasize theoretical knowledge while lacking practical support.However, “AI Embedded Systems: Algorithm Optimization and Implementation” integrates machine learning algorithm optimization theory and application practice, not only targeting embedded applications of artificial intelligence but also covering multi-dimensional machine learning algorithm optimization theories and techniques based on computation volume, memory, power consumption, etc. It elaborates on aspects from the underlying units of algorithms to joint optimization of machine learning algorithm training-inference and automatic optimization deployment algorithms, and also validates them with general embedded processors.The practical value of this book can also be seen from the selection of processor system architecture and machine learning algorithms.Now, many machine learning algorithms are based on massive storage and computation, with high requirements for power consumption, size, computing capacity, and storage capacity. Different fields of embedded systems have multiple paths, and how to choose in specific scenarios increases the difficulty of learning for developers to a certain extent.For example, GPU-based general-purpose multiprocessor architectures have strong versatility and high computing power, but the cost of hardware is high and power consumption is large; customized solutions based on dedicated computing acceleration engines have high computational efficiency and controllable power consumption, but the computational structure is relatively fixed and flexible; while solutions that deeply optimize algorithms based on existing processors are low-cost and do not require specialized or customized hardware, but the trade-off is that each case needs to be manually optimized, making development difficult.Existence is reasonable; different paths have their value in different scenarios.“AI Embedded Systems: Algorithm Optimization and Implementation” focuses on achieving machine learning algorithms through different levels of optimization based on existing embedded processor systems and architectures. Although the methods introduced are exemplified with general embedded processors, they can also be applied to GPU or hardware acceleration engine architectures. For instance, fast convolution algorithms based on transform domains can be applied in the underlying computation libraries of GPUs, and constant multiplication operations based on addition-subtraction graphs can be easily implemented in chip hardware design.Diversity exists not only in processor systems and architectures but also in the selection of algorithms in machine learning across multiple different fields.For example, algorithms based on statistical theory are designed with exquisite theoretical models, have good generalization, and relatively low computation volume, making them easy to implement in embedded systems with limited storage. However, they rely on strong probabilistic assumptions, and the bias in the model limits their performance in practical applications.In contrast, neural network algorithms, despite their superior performance, have large parameter storage requirements that limit their implementation in embedded systems. Recent research has found that many neural network architectures contain a lot of redundant computations and parameters, and simplification of computations and pruning of architectures can greatly reduce the computational performance and memory demands of neural networks, making it possible to implement them in embedded systems.Therefore, the content discussed in “AI Embedded Systems: Algorithm Optimization and Implementation” also covers the implementations of machine learning algorithms based on statistics and neural networks in embedded systems, introducing specific optimization methods through algorithms and examples.“Small Thoughts” Help Readers Absorb Knowledge Points FasterWhy can “AI Embedded Systems: Algorithm Optimization and Implementation” meet the needs of both developers and professional students?
First, regarding professionalism, both authors of this book have strong professional backgrounds. For example, Ying Rendong is an associate professor and master’s supervisor at the School of Electronic Information and Electrical Engineering at Shanghai Jiao Tong University, and has long engaged in teaching and research in embedded systems and digital circuits; Liu Peilin is a professor and doctoral supervisor at the same school. They possess both a professional perspective and an understanding of how to teach and clarify doubts.Secondly, regarding the thoroughness of the content, this book covers general embedded optimization technologies, including optimizations based on SMD instruction sets, memory access patterns, parameter quantization, etc., and introduces optimizations at the signal processing level, AI inference algorithm optimization, and optimization theories and methods for AI algorithms based on neural networks. Furthermore, it introduces basic knowledge about the development of automatic optimization and deployment tools for machine learning algorithms in general embedded environments through multiple examples that automatically search for optimization parameters and generate C code, aiming to enable readers to master the knowledge and skills for efficient AI embedded system development through practice based on understanding algorithms.Third, “AI Embedded Systems: Algorithm Optimization and Implementation” also considers the acceptability of different audiences. Therefore, the content on machine learning algorithms in embedded systems provides not only manual optimization steps but also software-assisted optimization introductions, allowing readers to understand based on their actual needs. In particular, the underlying algorithm codes in the book are presented in Python language. This is due to Python’s expressive capability and program simplicity. These “small thoughts” can help readers absorb and learn the knowledge points of this book better.With the widespread adoption of artificial intelligence technology across various industries, combining AI technology with embedded systems to build AI embedded systems lays the foundation for the large-scale commercialization of AI, which is also the original intention of the book “AI Embedded Systems: Algorithm Optimization and Implementation”.ClickRead the original text to learn more about “AI Embedded Systems: Algorithm Optimization and Implementation”
