New Book Recommendations | Embedded Artificial Intelligence

《Embedded Artificial Intelligence》

Compiled by Li Bin

ISBN：978-7-302-62796-8

Price: 69.00 yuan

Scan the QR code to purchase books with discount

The advent of ChatGPT has revealed the prototype of general artificial intelligence (AGI) to mankind. People have discovered that when artificial neural networks reach a certain scale, coupled with anthropomorphic learning methods (such as RLHF), miracles suddenly occur, and intelligence begins to get rid of the shackles of the human brain and is realized on computers!

However, there is still a challenge that needs to be solved: Can we achieve the above miracles on machines that are comparable in size and power consumption to humans, such as humanoid robots?

Currently, ChatGPT requires the use of more than 30,000 GPUs for training and inference, with a total power of more than 10 million watts and a daily electricity bill of approximately US$50,000. The volume of the human brain is only about 1.5 liters and the power is less than 20 watts. If we want to embed general artificial intelligence into robots, or even smaller drones, mobile phones, smart home appliances or IoT devices, we will have to overcome the above challenges. This is the topic that the book "Embedded Artificial Intelligence" wants to explore.

In the field of embedded artificial intelligence, we will use higher energy efficiency to execute more streamlined algorithms and place them in a machine with smaller size and weight to complete real-time reasoning tasks in the real world. . In a word, we must implement artificial intelligence in the most streamlined way!

To overcome this challenge, the problem needs to be addressed at all levels. To sum up, to realize embedded artificial intelligence, five components are needed.

Embedded AI chip. It is an AI accelerator with higher energy efficiency.

Lightweight AI algorithm. It has lower computational complexity and fewer parameters, but its accuracy is comparable to large AI algorithms.

Model compression. It further trims redundant parameters in lightweight AI algorithms and expresses them in a more streamlined way.

Compilation optimization. It translates the model into an encoding more suitable for the instructions of the AI ​​accelerator.

Multi-level cascading application framework. It uses more appropriate algorithms at more appropriate times, thereby reducing overall system cost and energy consumption.

When these five components are used in combination, general artificial intelligence is expected to be truly embedded in the machine, allowing it to see, hear, speak, and think, and truly "live" and become another form of life. ——Artificial life.

Of course, there is a long way to go to realize the above dreams. This book is an introductory guide that explores the problems and outlines current methods and tools for solving them. It's not the final answer, but it provides the key to getting there.

If you are interested in realizing general artificial intelligence on robots, drones, mobile phones, smart home appliances, and Internet of Things devices, it is undoubtedly a useful introductory book.

Features of this book

1. Domestic The first comprehensive introduction to embedded artificial intelligence.

2. The content is rich and detailed, covering the principles, platforms and practices of embedded artificial intelligence.

3. Make a systematic summary of the principles of embedded artificial intelligence, the concept is novel and the organization is clear.

4. Comprehensive introduction to the implementation of embedded artificial intelligence platform, including mainstream embedded neural network chips and software frameworks.

5. Outline the development process of embedded artificial intelligence.

6. Combining principles with practice, combine code to demonstrate practical application cases of embedded artificial intelligence.

Application Example

In the last chapter of the book, we try to realize a flying sun umbrella based on a drone. This kind of umbrella can track people's movements in real time and block people from the sun at any time, thereby freeing people's hands and allowing you to enjoy the hot summer. You can happily go shopping, play, and become the focus of people's attention.

In order to enable unmanned parachutes to track human bodies, we will use machine vision methods. The human body is photographed by the camera on the front of the parachute, and the outline of the human body is calculated through a lightweight human posture estimation algorithm to obtain the positions of the person's head, facial features, and limb joints, and then analyze the movements of the human body to predict which direction the person will move forward. , retreat, turn or go up or downhill, so that the unmanned parachute can track this movement, adjust the spatial position of the unmanned parachute at any time, and achieve the following and sunshading effect on the human body.

Considering weight, power consumption and other reasons, a dedicated embedded AI chip must be used. In the book, NVIDIA Jetson is used as an example for development. Now, chips with stronger inference performance per watt, such as ARM-based SoCs, have been widely used to meet the challenges brought about by the rapid development of technology. But the processes and methods introduced in the book are still applicable.

In order for the above lightweight algorithm to meet the requirements of real-time reasoning on the embedded AI chip, it is necessary to use model compression and optimization technology to perform operations such as pruning and quantification on the model, reduce the parameters of the model, and allow reasoning Performance increases exponentially, power consumption decreases exponentially, and is optimized for dedicated embedded AI chip inference. The book uses TensorRT as an example to illustrate that if an ARM-based SoC is used, compression and optimization technologies such as Tensorflow Lite can be used.

Finally, with the help of the knowledge in this book, we can finally make the unmanned parasol fly!

Table of contents

Swipe up to read

Part 1 Principles

Chapter 1 Artificial Intelligence and Artificial Neural Network

1.1 What is artificial intelligence

1.2 What is artificial neural network

1.3 Human Brain

1.4 Basic composition of artificial neural network

1.4.1 Artificial Neuron

1.4.2 Structure of Artificial Neural Network

1.5 Learning mechanism of artificial neural network

1.6 Types of Artificial Neural Networks

1.7 Advantages of artificial neural networks

1.8 Deep Neural Network

1.8.1 What is a deep neural network

18.2 Common deep neural networks

1.8.3 Convolutional Neural Network

1.9 Neural Network Architecture Search (NAS)

1.9.1 Search space

1.9.2 Reinforcement Learning Search

1.9.3 Progressive Search

1.9.4 Discrete search

1.10 Transfer Learning

1.10.1 What is transfer learning

1.10.2 Types of transfer learning

1.10.3 Advantages of transfer learning

1.10.4 Transfer learning method

1.10.5 Transfer Learning and Embedded Artificial Intelligence

Chapter 2 Embedded Artificial Intelligence

2.1 What is embedded artificial intelligence

2.2 Why embedded artificial intelligence is needed

2.3 Initial attempt: cloud computing model

2.4 From cloud to device: local mode

2.5 Technical Challenges of Embedded Artificial Intelligence

2.5.1 Model scale

2.5.2 Energy efficiency

2.5.3 Memory access

2.5.4 Inference speed

2.5.5 Dimensions and weight

2.6 Implementation approaches of embedded artificial intelligence

2.7 Implementation components of embedded artificial intelligence

Chapter 3 Principles of Embedded AI Chips

3.1 Parallel Computing

3.2 Pulsating Array

3.3 Multi-level cache

3.4 Data flow

Chapter 4 Lightweight Neural Network

4.1 Reduce computational complexity

4.1.1 Grouped convolution

4.1.2 Depth direction convolution

4.1.3 Pointwise convolution

4.1.4 Depthwise separable convolution

4.1.5 Channel out-of-order mixing

4.2 SqueezeNet

4.2.1 Core Idea

4.2.2 Network structure

4.2.3 Performance

4.3 Xception

4.3.1 Core Idea

4.3.2 Network structure

4.3.3 Performance

4.4 MobileNet v1

4.4.1 Core Idea

4.4.2 Network structure

4.4.3 Performance

4.5 MobileNet v2

4.5.1 Core Idea

4.5.2 Network structure

4.5.3 Performance

4.6 MnasNet

4.6.1 Core Idea

4.6.2 Network structure

4.6.3 Performance

4.7 MobileNet v3

4.7.1 Core Idea

4.7.2 Network structure

4.7.3 Performance

4.6 Application of lightweight neural network

Chapter 5 Deep Neural Network Compression

5.1 General method of neural network compression

5.1.1 Pruning

5.1.2 Weight Sharing

5.1.3 Quantification

5.1.4 Binary/Ternary value

5.1.5 Winograd convolution

5.2 Compression-compilation co-design

5.2.1 The concept of compression and compilation co-design

5.2.2 Compressor

5.2.5 Compiler

5.2.6 Advantages of compression and compilation co-design

Chapter 6 Embedded Neural Network Application Framework

6.1 The composition of hierarchical cascading system

6.2 Efficiency of hierarchical cascading system

6.4 Local-cloud collaboration mode

Chapter 7 Lifelong Deep Learning

7.1 Defects and reasons of traditional deep learning

7.2 The goal of lifelong deep learning

7.3 Characteristics of lifelong deep learning

7.4 Neurobiological Implications

7.5 Implementation of lifelong deep neural network

7.5.1 Dual learning system

7.5.2 Real-time update

7.5.3 Memory merging

7.5.4 Adapt to real scenarios

7.6 Lifelong Deep Learning and Embedded Artificial Intelligence

Part 2 Platform

Chapter 8 Embedded Neural Network Hardware Accelerator

8.1 Overview

8.2 NVIDIA Jetson

8.2.1 Jetson module introduction

8.2.1 Jetson module internal structure

8.2.3 Jetson performance

8.3 Intel Movidius

8.3.1 Movidius Myriad X VPU chip

8.3.2 Intel Neural Compute Stick

8.4 Google Edge TPU

8.4.1 Introduction to Google Edge TPU

8.4.2 Working principle of Google Edge TPU

8.5 XILINX DPU

8.6 ARM Ethos NPU

8.6.1 ARM Machine Learning Processor

8.6.2 Ethos-N Series

8.6.3 Ethos-U Series

Summary

Chapter 9 Embedded Neural Network Software Framework

9.1 Tensorflow Lite

9.1.1 Introduction to TensoFlow Lite

9.1.2 How TensorFlow Lite works

9.2 TensorRT

9.3 OpenVINO

9.3.1 Introduction to OpenVINO

9.3.2 The composition of Open VINO

9.3.3 Open VINO application development

9.4 XILINX Vitis

9.5 uTensor

9.6 Apache TVM

Summary

Part 3 Realization

Chapter 10 Building an Embedded Neural Network Development Environment

10.1 Embedded AI development process

10.2 NVIDIA Jetson development process

Chapter 11 Optimizing Embedded Neural Network Model

11.1 TensorFlow model optimization

11.1.1 Post-training optimization

11.1.2 Optimization during training

11.2 TensorRT model optimization

11.2.1 Integrate with mainstream deep learning frameworks

11.2.2 Deployment to embedded systems

11.2.3 TensorRT API

11.2.4 TensorRT application example

11.2.5 Model Converter

11.3 Comparison of two model optimization techniques

Chapter 12 Performing Inference on Embedded Devices

12.1 Building the project from source

12.2 Use ImageNet to implement image classification

12.2.1 Static image classification

12.2.2 Camera real-time video classification

12.3 Use DetectNet to implement target detection

13.3.1 Static image target detection

13.3.2 Camera real-time video target detection

12.4 Using SegNet to implement semantic segmentation

12.4.1 Static image semantic segmentation

12.4.2 Video semantic segmentation

12.5 Using PyTorch to implement transfer learning

12.6 Using transformed models

Chapter 13 Embedded Neural Network Application Example

13.1 Application scenarios

13.2 Hardware selection

13.3 Model Development

Conclusion: Intelligence of all things

The above is the detailed content of New Book Recommendations | Embedded Artificial Intelligence. For more information, please follow other related articles on the PHP Chinese website!