Big data is a new concept emerging with the explosive growth of information data and the rapid development of network computing technology. According to the definition of McKinsey Global Institute, big data is a collection of data whose scale is so large that its acquisition, storage, management, and analysis far exceed the capabilities of traditional database software tools. The four characteristics of low data type and value density. Big data can help enterprises in all walks of life to dig out the needs of users from the originally worthless massive data, so that the data can change from quantitative to qualitative, and truly generate value. With the development of big data, its application has penetrated into various fields such as agriculture, industry, commerce, service industry, and medical field, and has become an important factor affecting the development of the industry.
The artificial intelligence that people currently call refers to a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence, and is manifested by artificially manufactured systems. of intelligence.
Traditional artificial intelligence is limited by computing power, and has not been able to complete large-scale parallel computing and parallel processing, and the ability of artificial intelligence systems is poor. In 2006, Professor Hinton proposed "deep learning" neural network to make breakthroughs in artificial intelligence performance, which in turn prompted the artificial intelligence industry to enter a stage of rapid development again. The main mechanism of the "deep learning" neural network is to simulate the learning process of the human brain through the deep neural network algorithm, and combine the low-level features into a higher-level abstract representation through the nonlinear relationship between input and output, and finally achieve the level of mastery and application. The richness of the amount of data determines whether there is enough data to train the neural network, so that the artificial intelligence system can reach the level of strong artificial intelligence after deep learning training. Therefore, whether there is enough data for in-depth training of artificial neural networks and improving the effectiveness of the algorithm is one of the determinants of whether artificial intelligence can reach the human-like or superhuman level.
With the explosion of the mobile Internet, the amount of data has grown exponentially, and the accumulation of big data has provided basic support for artificial intelligence. At the same time, benefiting from the breakthrough of computer technology in data collection, storage, calculation and other aspects, artificial intelligence has evolved from a simple algorithm + database development to a state of machine learning + deep understanding.
2. Artificial Intelligence Industry and Ecology
According to the industrial chain structure, artificial intelligence can be divided into basic technology layer, AI technology layer and AI application layer. The basic technology layer mainly focuses on data resources, computing capabilities and hardware platforms. Data resources are mainly all kinds of big data, and hardware resources include chip research and development, storage device development, etc. The AI technology layer focuses on algorithms, models and applicable technologies, such as computational intelligence algorithms, perceptual intelligence algorithms, and cognitive intelligence algorithms. The AI application layer mainly focuses on combining artificial intelligence with various downstream fields, such as drones, robots, virtual customer service, and voice input methods.
Figure 1 Artificial intelligence industry chain
Source: China Industry Information Network, "2017 China Artificial Intelligence Industry Development Overview and Future Development Trend Analysis"
(1) Basic technology layer
1.1 Big Data
Data resources are the basic material for machine learning training. Through data learning, machines can continuously accumulate experience and optimize decision-making parameters, gradually becoming closer to human intelligence.
From the perspective of data flow direction, the industrial chain of big data can be divided into three levels: underlying platform, processing and analysis, and application. The underlying platform is composed of infrastructure and data asset pools, mainly providing data collection, sharing and transaction services. Processing and analysis is based on the original data, after cleaning the data and presenting it in different ways. On the basis of data processing and analysis, it mines the data needs of various industries, and finally provides services to users.
According to the degree of data application, the functions of each participant in the big data industry chain can be subdivided into seven aspects: data standardization and normalization, data collection, data security, data storage and management, data analysis and mining, data operation and maintenance, and data utilization.
1.2 Computing power and hardware platform
Data resources, core algorithms, and computing power are the three core elements of artificial intelligence. With the rapid development of the global mobile Internet and the Internet of Things, the data available to humans is exploding. Massive big data will bring invaluable value to the development and application of artificial intelligence through the latest deep learning technology, and the improvement of computing power is the premise guarantee for the development of artificial intelligence. Among them, the chip is the core of computing power.
At present, there are four main types of AI chips: GPU, FPGA, ASIC and human-like chips.
1.2.1 GPU
1.2.1.1 Introduction to GPU
GPU is a graphics processor, which was originally a microprocessor used for image operations. The GPU optimizes and adjusts the CPU structure, making it faster and more powerful to handle floating-point operations. In 2009, Andrew Ng and his team at Stanford University discovered that GPU chips could run neural networks in parallel. Using GPUs to run machine learning models, the same large training set, GPUs can support 10-100 times more application throughput than when using CPUs alone with less power consumption and less infrastructure. So GPUs have become the processors for data scientists to process big data.
1.2.1.2 Status Quo of GPU Industry
At present, the international GPU market is divided between NVIDIA and AMD. More than 70% of the global GPU industry market share is occupied by NVIDIA, while the general-purpose computing GPU market used in artificial intelligence is basically monopolized by NVIDIA. At present, the company has established and cooperated with Google, Microsoft, IBM, Toyota, Baidu and many other companies that try to use deep neural networks to solve massive and complex computing problems. NVIDIA has deepened its cooperation with downstream customers in the field of deep learning, and has developed a number of GPU products for deep learning. From the perspective of product maturity and the scale of the ecosystem, NVIDIA's GPUs already have a dominant position.
China started relatively late in the field of GPU chip design. At present, only two companies, Jingjiawei and Zhaoxin, which master the core technology, are gradually breaking the monopoly of foreign chips in my country's GPU market, but the products are still mainly used for the initial graphics display control of GPUs. In the field, there is still a long way to go from the GPU technology required by artificial intelligence.
1.2.2 FPGA
1.2.2.1 Introduction to FPGA
FPGA, the field-effect programmable logic gate array, was originally a semi-custom programmable circuit developed from an application-specific integrated circuit. It is modified by programming like software. Different programming data can generate different circuit functions on the same FPGA, which has strong flexibility and adaptability.
Both FPGAs and GPUs have a large number of computing units within them, so they are both powerful in computing power. When performing neural network operations, the speed of the two will be much faster than the CPU. However, due to the fixed architecture of the GPU, the instructions natively supported by the hardware are also fixed, while the FPGA is programmable. Its programmability is key, as it gives software and end-application companies the ability to offer solutions that differ from their competitors and the flexibility to modify circuits for the algorithms they use. Compared with GPU, FPGA has the characteristics of high performance, low power consumption and hardware programmable.
1.2.2.2 Status Quo of FPGA Industry
At present, the entire FPGA market is occupied by two foreign giants. According to data from the Orient Securities Research Institute, Xilinx and Altera account for nearly 90% of the total, with a total of more than 6,000 patents. The remaining shares are occupied by Lattice and Microsemi. There are more than 3000 patents in total. The limitation of technology patents and the long development cycle make the FPGA industry have extremely high barriers.
Although the Chinese government has invested tens of billions of scientific research funds in this field for many years, the patent restrictions and technical thresholds of FPGA make the research and development of China's FPGA very difficult. Some progress has been made in the research and development of FPGA, but there is a big gap between the product performance, power consumption, capacity and application field compared with foreign advanced technology. At present, some domestic capitals have tried to go abroad and enter the FPGA industry by merging semiconductor companies to achieve overtaking in corners.
1.2.3 ASIC
1.2.3.1 Introduction to ASIC
ASIC, that is, an application-specific integrated circuit, refers to an integrated circuit designed and manufactured in response to the requirements of a specific user or the needs of a specific electronic system. As a product of integrated circuit technology and a specific user's complete machine or system technology, ASIC has the following advantages compared with general integrated circuits: smaller size, lower power consumption, improved reliability, and improved performance , Confidentiality is enhanced. FPGAs are generally slower than ASICs, cannot complete more complex designs, and consume more power, so ASICs are far superior to FPGAs in terms of power; however, the specialized features of ASICs make them expensive to produce. If the shipments are small, it is less economical to adopt ASICs. Once the artificial intelligence technology matures, the characteristics of ASIC-specific integration will achieve scale effects, and the cost will be greatly reduced compared with general-purpose integrated circuits.
At present, there are not many applications of ASIC in artificial intelligence deep learning, but we can do similar reasoning with the development of bitcoin mining machine chips. Bitcoin mining and artificial intelligence deep learning are similar in that they both rely on the underlying chip for large-scale parallel computing. The chips of Bitcoin miners go through four stages: CPU, GPU, FPGA and ASIC. Among them, ASIC has shown unique advantages in the field of Bitcoin mining. With more and more applications of artificial intelligence in various fields and showing superior performance, ASIC is promising in the long run.
1.2.3.2 Current Situation of ASIC Market
With the rise of artificial intelligence, technology giants have deployed chip manufacturing. Qualcomm, AMD, ARM, Intel and NVIDIA are all working on integrating custom chips into their existing solutions. Nervana and Movidius (both currently owned by Intel) are said to be working on collective solutions. The more mature product in ASIC is the TPU developed by Google for AlphaGo. The first-generation TPU product was officially launched by Google at the 2016 I/O conference. At the developer I/O conference in May this year, Google officially announced the second-generation TPU, also known as Cloud TPU. Compared with the first-generation TPU, It can be used for both training neural networks and inference, and the floating-point performance is 15 times higher than that of traditional GPUs.
The application of ASIC in the field of artificial intelligence started late, and the level at home and abroad is not much different. At present, there are several domestic companies dedicated to the research of artificial intelligence-related ASIC chips, and the representative companies are Horizon Robot, Zhongke Cambrian and Zhongxing Microelectronics. Among them, Horizon Robotics, as a start-up company, is committed to building an artificial intelligence "brain" platform based on deep neural networks - including software and chips, which can solve problems such as environmental perception, human-computer interaction, and decision-making control with low power consumption and localization. . Its research and development on the chip is still immature. Zhongke Cambrian and Zhongxing Microelectronics already have relatively mature products. Cambrian chips are specially oriented to deep learning technology, and have developed the world's first dedicated processor chip NPU for deep learning. The three chips that have been developed so far are respectively oriented towards the prototype processor structure of neural networks, large-scale neural networks and various machine learning algorithms. , It is expected that the industrialization of the chip will be realized in 2018. Zhongxing Microelectronics launched China's first embedded neural network processor (NPU) chip in June 2016, which is the world's first embedded video capture and compression coding system-on-chip with deep learning artificial intelligence. This deep learning-based chip is used in face recognition, with a maximum accuracy rate of 98%, exceeding the recognition rate of the human eye. The chip was mass-produced on March 6, 2017, and as of May this year, shipments were over 100,000 units.
1.2.4 Brain-like chips
1.2.4.1 Introduction to Human Brain Chip
The human-like brain chip is a new type of ultra-low power consumption based on neuromorphic engineering and learning from the information processing method of the human brain. computing chip. Theoretically, the human brain-like chip is closer to the artificial intelligence target chip, trying to imitate the working principle of the human brain in the basic architecture, using neurons and synapses to replace the traditional architecture system, so that the chip can be asynchronous and parallel. , vulgar and distributed ability to process information and data, and at the same time have the ability of self-protection perception, recognition and learning.
1.2.4.2 Market Status of Human Brain Chips
Brain-like chips are the key direction for the development of artificial intelligence chips. At present, governments and technology giants of various countries are vigorously promoting the research and development process of human-like brain chips. Developed countries, including the United States, Japan, Germany, the United Kingdom, and Switzerland, have formulated corresponding development strategies. China's human-like brain research projects have also been officially launched start up. At present, a number of technology companies in the world are at the forefront and have made breakthroughs in the research and development of human-like brain chips. Representative products include IBM's TrueNorth chip, Qualcomm's Zeroth chip, and Google's "Neural Network Turing Machine".
(2) AI technology layer
The AI technology layer mainly focuses on algorithms, models and applicable technologies. According to the different levels of intelligence, artificial intelligence can be divided into three stages: computational intelligence, perceptual intelligence, and cognitive intelligence. Computational intelligence, that is, fast computing and memory storage capabilities, at this stage is mainly the combination of algorithms and databases, so that machines can begin to calculate and transmit information like humans; Perceptual intelligence, that is, visual, auditory, tactile and other perceptual capabilities, in this In the first stage, the combination of the database and the shallow learning algorithm enables the machine to begin to understand and understand, and to make judgments and take actions; cognitive intelligence, that is, the ability to understand and think, this stage mainly uses deep learning algorithms, Enables machines to think like humans and take the initiative to act.
The AI technology layer can be divided into a framework layer and an algorithm layer. The framework layer refers to frameworks or operating systems such as TensorFlow, Caffe, Theano, Torch, DMTK, DTPAR, and ROS, and the algorithm layer refers to the data processing method.
Depending on the type of data, a problem is modeled in different ways, that is, learning methods. According to the classification of learning methods, artificial intelligence algorithms can be divided into traditional machine learning and neural network algorithms, of which traditional machine learning can be further subdivided into supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning.
2.1 Traditional Machine Learning
2.1.1 Supervised Learning
Under supervised learning, the input data is called "training data", and each set of training data has a clear identification or result, such as "spam" and "non-spam" in the anti-spam system, and for handwritten digit recognition. "1", "2", "3", "4" etc. When building a prediction model, supervised learning establishes a learning process, compares the prediction result with the actual result of the "training data", and continuously adjusts the prediction model until the prediction result of the model reaches an expected accuracy rate. Common application scenarios of supervised learning are classification problems and regression problems. Commonly used algorithms include regression algorithm, Naive Bayes, SVM, etc.
2.1.2 Unsupervised Learning
In unsupervised learning, the data is not specifically identified, and the model is learned to infer some intrinsic structure of the data. Common application scenarios include the learning of association rules and clustering. The common algorithms for association rule learning are the Apriori algorithm and its extension algorithms, and the commonly used algorithms for clustering include the k-Means algorithm and its similar algorithms.
2.1.3 Semi-Supervised Learning
In this learning method, the input data is partially marked and partially unmarked. This learning model can be used to make predictions, but the model first needs to learn the internal structure of the data in order to organize the data reasonably to make predictions. Application scenarios include classification and regression, and algorithms include some extensions to commonly used supervised learning algorithms that first attempt to model unlabeled data and then make predictions on labeled data. Graph Inference or Laplacian SVM.
2.1.4 Reinforcement Learning
In this learning mode, the input data is used as feedback to the model. Unlike the supervised model, the input data is only used as a way to check whether the model is right or wrong. In reinforcement learning, the input data is directly fed back to the model, and the model must immediately respond to this. make adjustments. Common application scenarios include dynamic systems and robot control. Common algorithms include Q-Learning and Temporal difference learning.
2.2 Neural Network
Artificial neural network is a simulated biological neural network, which is formed by connecting a large number of neurons with adjustable connection weights. It has the characteristics of large-scale parallel processing, distributed information storage, and good organizational learning ability. It learns through certain learning criteria. Then establish relevant models to solve certain tasks. In the aspect of learning algorithm design of artificial neural network, the artificial neural network is generally trained and adjusted with a large amount of data, and the parameters of nodes at all levels are constantly revised. Through continuous learning, the artificial neural network has preliminary adaptive ability and self-organization ability and strong Based on its many advantages, artificial neural networks have become the core of artificial intelligence algorithms. The deep learning algorithm is the latest algorithm of artificial neural network. Its essence is to learn more useful features through many hidden layer machine learning models and massive training data, thereby improving the accuracy of classification or prediction.
(3) AI application layer
The application of artificial intelligence mainly adopts the method of "AI + vertical industry" to penetrate into traditional industries. According to the different levels of development, it can be divided into three levels: dedicated artificial intelligence, general artificial intelligence and super artificial intelligence. Among them, dedicated artificial intelligence is mainly based on one or more specialized fields and functions; general artificial intelligence means that machines have the same possibility of doing all work as humans, and the key lies in automatic cognition and expansion; super artificial intelligence refers to self-awareness , including independent values, world views, etc., currently only exist in the concept of cultural works.
According to the type of application technology, the application technology of artificial intelligence can be divided into four parts: computer vision, machine learning, natural language processing and robotics.
3.1 Computer Vision
Computer vision refers to the ability of computers to recognize objects, scenes and activities from images. Computer vision technology uses a sequence of image processing operations and other techniques to decompose image analysis tasks into small tasks that are easy to manage. At present, computer vision is mainly used in face recognition and image recognition (including static and dynamic two types of information) .
Face recognition, also known as portrait recognition and facial recognition, is a biometric recognition technology based on human facial feature information. A series of related technologies that use cameras or cameras to capture images or video streams containing faces, and automatically detect and track faces in the images, and then process the detected faces.
Image recognition is a technology in which computers process, analyze and understand images to identify targets and objects in various patterns. The recognition process includes image preprocessing, image segmentation, feature extraction and judgment matching. Due to the technical limitations of dynamic monitoring and recognition, the research on static image recognition and face recognition is temporarily in a leading position.
At present, foreign technology giants deploy their own research and development and acquisitions in the field of computer vision, widely use technology to upgrade their own products, and build technical service platforms and new categories based on their own genes to continuously increase their influence. All domestic BATs in China have already deployed related fields and conducted functional research and development based on their own products. Baidu is relatively more aggressive, setting up an independent venture capital firm focused on early AI investments.
In addition to the BAT three giants, there are also many domestic start-up companies involved in computer vision technology, mainly focusing on technology applications. One of the typical representatives is Megvii Technology. The company was established in November 2012. The company focuses on the application research of face recognition technology and related products, and provides services for developers. It can provide a complete set of visual technology for face detection, face recognition, face analysis and face 3D technology. Service, mainly by providing cloud API, offline SDK, and user-oriented independent research and development product form, the face recognition technology is widely used in Internet and mobile application scenarios. Face++ has cooperated with many Internet companies and mastered a database of 5 million face pictures through "desensitization" technology. The accuracy rate of face recognition LFW in Internet pictures has reached 99.6%. The partners include Ali, 360 and a number of large-scale Image, social, and device businesses.
At present, the popularity of computer vision entrepreneurship in China continues to increase. According to iiMedia Research (iiMedia Consulting) data, among the field distribution of Chinese artificial intelligence startups, the computer vision field has the most startups, up to 35.
3.2 Machine Learning
Machine learning refers to the ability of computers to make predictions and make optimal decisions by processing, analyzing and learning a large amount of existing data. At its core, machine learning is the automatic discovery of patterns from data, which can be used to make predictions once discovered.
Machine learning has a wide range of applications and has the potential to improve the performance of almost everything that generates huge amounts of data. In addition to fraud detection, these activities include sales forecasting, inventory management, oil and gas exploration, and public health. Machine learning technology also plays an important role in other cognitive technology fields, such as computer vision, which can improve its ability to recognize objects by continuously training and improving visual models in massive images.
Today, machine learning is one of the hottest research areas in cognitive technologies, attracting nearly a billion dollars in venture capital during the period 2011-2014. Google also spent $400 million to acquire Deepmind, a company that researches machine learning technology, in 2014. At present, the number of domestic machine learning related enterprises is relatively small. BAT has inherent advantages in machine learning, and the fourth paradigm of domestic startups is a solution provider based on machine learning.
3.3 Natural Language Processing
Natural language processing is to use artificial intelligence to process, understand and use human language, and to predict the probability distribution of language expressions by establishing language models, so as to achieve the goal.
Natural language processing technology is widely used in life, such as machine translation, handwritten and printed character recognition, text conversion after speech recognition, information retrieval, extraction and filtering, text classification and clustering, public opinion analysis and opinion mining. They apply technologies such as syntax analysis, semantic analysis, and text comprehension in natural language processing, and they are the most cutting-edge research areas in the field of artificial intelligence. Today, the development of AI in these technical fields has increased the recognition accuracy rate from 70% to more than 90%, but only when the accuracy rate increases to 99% and above can it be recognized that the technology of natural language processing reaches human level.
With the help of capital and industry, my country's artificial intelligence speech recognition technology has been at the international leading level, and the technology is mature. In terms of general recognition rate, all enterprises have maintained a level of about 95%. Listed companies such as Baidu and iFLYTEK occupy the forefront of the market by virtue of their profound technology and data accumulation, and continue to evolve their service capabilities through the development of software and hardware services. Yunzhisheng, which released the second "public cloud for speech recognition" in China after iFlytek, also occupies a large market space in the provision of various general voice service technologies. In addition, Zidong Ruiyi and Naxiang Cube, which rely on the Institute of Automation of the Chinese Academy of Sciences, and Suzhou Spiech, which has an overseas background, occupy a leading position in speech recognition in the field of education.
3.4 Robots
The integration of cognitive technologies such as machine vision and autonomous planning into extremely small but high-performance sensors, actuators, and cleverly designed hardware has given rise to a new generation of robots capable of working alongside humans and capable of Flexibility to handle different tasks in an unknown environment.
At present, at least 48 countries in the world are developing robots, of which 25 countries have been involved in the development of service robots. In Japan, North America and Europe, so far, more than 40 service robots of 7 types have entered the field of experimental and semi-commercial applications in the field of service robots. The United States is the birthplace of robots, and its robotics technology has always been in a leading position in the world. Its technology is comprehensive, advanced, and adaptable. It has an absolute advantage in the military, medical, and home service robot industries, accounting for the service robot market. about 60% share. The research and development of the domestic intelligent robot industry mainly focuses on three aspects: home robots, industrial/enterprise services and intelligent assistants. Among them, industrial and enterprise service robot R&D enterprises are in a relatively leading development stage relying on the policy background and market demand. However, among the enterprises involved in intelligent robots in China, enterprises engaged in the research and development of home robots and intelligent assistants account for the vast majority.
因为服务一般都要结合特定市场进行开发，本土企业更容易结合特定的环境和文化进行开发占据良好的市场定位，从而保持一定的竞争优势；另一方面，外国的服务机器人公司也属于新兴产业，大部分成立的时候还比较短，因而我国的服务机器人产业面临着比较大的机遇和可发展空间。