Shulou(Shulou.com)06/03 Report--

Looking back at 2019, the development of AI technology is changing the trend of science and technology in the world.

In 2019, AI, as one of the technological cores of the "fourth Industrial Revolution", was still a global hot spot in science and technology, and seemed to be getting hotter.

You see, in the past NeurlPS2019, 13000 people attended the conference and 1428 papers were published, setting a new record. The AI field is full of talents and the ecology is prosperous.

This year's AI development direction, it can be said, more deep into a variety of industries, but also closer to people's clothing, food, housing and transportation. You see, all kinds of hot searches on Weibo have something to do with AI.

Today, through 5 hot searches, I will show you the hot spots of global AI development this year.

01 # Google realizes Quantum hegemony # look at the improvement of AI's computing power

On October 23, with the 150th anniversary issue of Nature magazine, Google announced the realization of "quantum hegemony."

Google used a 54-qubit quantum processor, Sycamore, to do what the world's most powerful supercomputer, Summit, took 10, 000 years to do in 200 seconds-random sampling of quantum circuits.

As soon as the news was announced, it immediately became the biggest hot spot of science and technology in the world.

The industrial development of AI has three elements: computing power, algorithms and data, and "quantum computing" will unlock the new world of human computing power and open up the possibility of new dimensions for AI.

As a result, Google began laying out quantum computing 13 years ago, and later set up a GoogleAI Quantum team to explore.

Although the realization of quantum hegemony is still a "HelloWorld" computing task according to Pichai, it is enough to make 2019 an important year in the history of AI computing.

If "quantum computing" is still a long way off, there will also be a lot of progress in 2019 in terms of a more reliable improvement in computing power. The progress in the industry is mainly in two areas: chip research and development and the new cloud computing center.

In terms of chips, academia and industry are exploring more possibilities.

For example, it is more general.

On August 1st, Nature magazine introduced the "Tianji" AI chip architecture from the Chinese team. Led by Tsinghua University, the joint research of many Chinese and foreign universities, combined with the AI algorithm of brain-like computing, successfully drives the self-driving car. This achievement shows the possibility of realizing the ultimate "universal" AI.

And bigger ones.

Also in August, CerebrasSystems launched Wafer-Scale Engine (WSE) in Silicon Valley, the largest AI chip in history, which is 56 times larger than an ordinary chip and has 1.2 trillion transistors. Then, a new deep learning computer CS-1 with WSE was launched. Such supercomputers provide new computing power for large-scale AI problems, and the training time will be greatly shortened.

And, more focused.

In November, after Nvidia and Qualcomm launched the artificial intelligence special chip (TPU), Intel finally launched the processor NervanaNNP, which solves complex deep learning computing problems. Baidu developed Kunlun, China's first full-featured cloud AI chip, and launched Kunlun Cloud servers in December, while supporting mainstream AI frameworks / platforms such as Flying Propeller.

On the edge device, it is more "specialized". In July, following the new design idea of "software-defined chip", Baidu released a Baidu swan chip specially built for far-field voice interaction, which is mainly used in car voice interaction, smart home and other scenes. It can be said that the configuration of dedicated AI chips has become a standard in both edge devices and servers.

As for the cloud computing center, the theme is to use AI to optimize the cloud computing center dedicated to AI.

With the increasing demand for AI and the increasing demand for computing power, the existing computer room has been unable to meet the computing needs. According to IDC, 55% of the world's corporate data centers will be forced to be updated or optimized by the end of 2020.

Refrigeration and energy conservation is an important part of it. According to Gartner analysis, by the end of 2020, as the demand density of AI computing increases, 30% of data centers will no longer have economic value.

At the beginning of this year, the Ministry of Industry and Information Technology, in conjunction with relevant departments, issued the guidance on strengthening the Construction of Green data Centers, clearly pointing out that by 2022, the energy consumption of China's data centers should reach the world's advanced level, and the designed energy use efficiency of large and super-large data centers should not be higher than 1.4.

China is also making good progress after DeepMind achieved remarkable results in cooling data centers with AI in 2018. Huawei continues to release innovative technologies such as iPower, iCooling and iManager, optimizing distribution systems and end-to-end energy efficiency management systems through AI, and increasing resource utilization by 20%.

The cloud computing center under construction by Baidu has super-large neural networks, super-large-scale high-performance computing clusters, and adopts the architecture of Baidu fourth-generation data center infrastructure module, based on the AI control system developed by Baidu Flying Propeller, to achieve an average annual PUE of 1.15, becoming the first fully distributed prefabricated case of power supply and cooling in the data center field.

In 2019, the global AI computing power continues to improve, chips are more powerful and more possible, and computing centers are more energy-efficient and more economical.

In the future, computing power will continue to be a key part of the layout breakthrough of the entire industry, and quantum computing will usher in a new round of outbreak. The end-side AI chip will be more low-cost, professional and integrated solution. At the same time, NPU will become the basic module of the next generation end-side general CPU chip.

02 # parking Robot in Daxing Airport # look at the landing of AI industrialization

On September 25, Daxing Airport was officially put into operation, making it the largest airport in the world and the most "high-tech" airport.

Among them, the automatic parking robot put into use, so that the eyes of the public bright, boarded the hot search. It is launched by JD.com Mathematical Science and Shougang S-PARK. When you drive into the designated parking area, you can directly get out of the car and leave, and the rest of the robots looking for parking spaces will do it for you.

In fact, this is the automatic guidance robot AGV, which has long been used in logistics, warehousing and other fields, which is also the representative of a large class of industrial robots.

The traditional industrial robot only relies on the electromechanical system and the traditional automatic control theory, which is suitable for accomplishing the given tasks on the production line. However, with the advance of global industry 4.0, the demand for more intelligent and flexible applications continues to emerge.

Relying on industrial big data and sensors, robots powered by AI can "see" and "understand" and begin to become smarter. You can not only handle more situations on your own, but also cooperate with others.

For example, in July, at the Baidu AI developer conference, Baidu unveiled the robotic arm "Dr. Tea" based on the service robot program. Using Baidu 3D vision technology, the robot can detect and track the position of the teacup. Through motion planning and control, the manipulator can carry out collision detection, avoid obstacles, and automatically generate tea pouring tracks. Voice and semantic interaction, so that the robot like a "human" in general, understand, speak well, do accurately.

In addition, the Boston power robot in the United States, which used to be a hot search on various robot action shows, has finally begun its industrial application this year. In April, after the acquisition of KinemaSystems, the company officially launched a commercial logistics robot, PICK. It can rely on identification technology to load and unload mixed SKU, which is almost equal to human beings. In July, the company announced the commercial use of the security quadruped robot SPOT.

It can be said that in 2019, the development of AI has a more solid industrial landing, and more in-depth integration with the production practice of various industries.

Not only in the secondary industry, there are industrial robots to help manufacturing and industrial improve efficiency, in agriculture, we also see all kinds of AI applications. For example, institutions such as China Agricultural University and JD.com 's "Shennong brain" can prevent and control diseases and insect pests and raise them intelligently; the intelligent soilless cultivation solution jointly created by Baidu and BOE Plant Factory can save energy and water, reduce costs and increase output.

These AI applications digitize and produce the personal experience of original agricultural experts, so that their business has the possibility of large-scale production.

In the future, the application of AI will be more in-depth, and various industries will apply deep learning technology on a large scale to innovate, accelerate transformation and intelligent upgrade.

03 # Shanghai will put AI trash can # look at AI universal platform

Garbage sorting has become a regular hot spot in China on July 1, as Shanghai began to implement the "strongest in history" regulation. "what kind of garbage is so-and-so" has become a buzzword.

A month later, the news that "2000 AI trash bins will be put into Shanghai" went viral. Each time a person puts a piece of garbage into the trash can, the system will automatically identify "what kind of garbage is this" and put it in the appropriate location.

Related applications, and Baidu APP's "AI garbage sorting" intelligent Mini Program, based on Baidu AI visual and voice capabilities, through voice search or picture recognition search to help users identify spam types.

In fact, the technology behind these AI applications is so mature that developers only need to call the basic AI algorithm and add a step of mapping. In fact, this can be achieved directly with the help of various AI frameworks and their preset models, and there may be no need to retrain data.

A small AI trash can application actually reflects the vast demand of the whole society for the basic capabilities of AI, while the rapid landing of thousands of similar needs in 2019 reflects the platform of the AI framework, the deepening of industrialization, and the enhancement of ease of use.

For example, from the progress of PyTorch and TensorFlow in 2019, we can see that both companies are making up for their shortcomings in order to light up more areas of the AI arena.

Because PyTorch itself is easy to use, it is almost Python-based code development, making it more popular in the academic community. But in the past, for reasons such as not supporting the deployment of edge devices, there were more use cases for TensorFlow, which has more deployment service capabilities, for industrial applications.

At this year's f8, PyTorch launched an experimental PyTorch Mobile. In version 1.3, full end-to-end workflow is supported from Python development to deployment on iOS and Android. TensorFlow, on the other hand, greatly enhanced the development experience at the World conference in October, incorporating more preset models, including its own Deepmind, and stressed that "Deploy ML anywhere" could be implemented through TensorFlow Extended, Lite and .JS.

Domestically, as the only open source and fully functional deep learning platform in China, the number of Baidu Flying Propeller developers has exceeded 1.5 million. This year's upgrade is moving towards standardization, automation and modularization to meet the increasing and complex needs of developers and industries. In the middle of the year, Feipao cooperated with Huawei Kirin chip to break through the deep learning framework and chip AI computing power to build China's AI core competitiveness.

In 2019, we see that the update strategies of the major AI frameworks are moving towards a more comprehensive and easy-to-use direction. In the future simple requirements development, technology may no longer be a problem, just use it. AutoML may soon be implemented, and as long as meta-knowledge is entered, the framework can automatically select data, automatically adjust the optimization model, and train and deploy.

With the improvement of the capability of the AI framework platform and the decline of the difficulty of development, in the future, there are likely to be several "AI factories" in the world, which will summarize and abstract the essence on the demand side, and then come up with a set of scalable solutions. For example, voice customer service and mass replication are applied to 1-1 education and financial services.

In the future, it may be more difficult to develop the demand and know how to land in the right posture, and it is also needed by the talents in the market.

04 # 5G official commercial # look at the smart city life is about to begin

On November 1, China's three major operators officially launched the 5G package, marking the official commercial use of the fifth generation mobile communication technology in China, and China has naturally become the world's largest 5G market.

As an infrastructure, 5G's new capabilities will work with AI to empower each other and open up more opportunities.

The "high concurrency" application scenario of 5G will enable more devices to connect to the network, and the Internet of everything will be truly realized.

According to IDC, global spending on the Internet of things is expected to be $726 billion this year and will break through the $1 trillion mark next year. This year, more smart hardware devices have emerged from personal wear to home use, and sales are growing at a new high.

Xiaomi, which has a wide range of smart hardware devices, from sockets to watches, has more than 210 million IoT platform device connections, up 62 per cent from a year earlier, according to its third-quarter results. In terms of smart speakers, the largest domestic shipments of small speakers achieved a year-on-year growth rate of 3700% in the second quarter.

As for developers, the capability of the platform continues to improve, and the growth rate of developers is even faster.

The Alexa platform released its annual summary in February, saying that it had more than 100000 compatible devices, 100m devices, 100 per cent improvement in user interaction and 15 more languages in terms of functionality. DuerOS releases 5. 0 systems, adding the industry's leading full-duplex no-wake-up capability. For developers, the number of smart hardware companies registered in China increased to 2027 from 1459 in 2018, an increase of nearly 40 per cent, according to Tianyan data.

The "low latency" application scenario of 5G opens up the landing of high reliability requirements such as unmanned cars and manufacturing.

With a large number of layoffs by major car companies in Japan, the United States and Europe this year, further commercial progress of unmanned vehicles in the United States has been slow. On the one hand, the cost is too high, companies are too short of money this year, and GM has postponed the deadline for its own self-driving taxis to hit the road. On the other hand, the development of self-driving cars for pure full roads is too difficult, and Waymo CEO John Krafcik also said in January that self-driving may never be able to drive in full road conditions.

In August, The Information released customer satisfaction figures for Waymo's unmanned taxis in Silicon Valley and Phoenix: 70 per cent. According to stringent taxi safety standards, this actually indicates that 30% of the cases, users have a perceived danger. For example, users complain that they are too close to the car, are not allowed to park or even park directly on the road.

China has made gratifying progress, with Baidu Apollo still taking the lead in road tests, with L4 self-driving city road test mileage reaching 3 million km. In terms of landing, with the release of ApolloEnterprise at the beginning of the year, Apollo also ushered in the first year of commercialization.

Small car OS has joined hands with more than 300 ecological partners, and the "Red Flag EV" Robotaxi team jointly developed by FAW Hongqi has begun trial operation in Changsha.

With the advent of 5G, the modified road surface of the smart city will provide more possibilities for the landing of unmanned cars in China.

In a narrow sense, smart city is to carry out various intelligent transformation of the city, adding street sensors, smart lampposts, garbage cans and even intelligent cool facilities. Relying on the low delay of 5G, these sensors can transmit road information to vehicles in real time. After the transformation, the road environment of the unmanned car will be greatly improved, and it will no longer be so difficult to land.

Google's smart city renovation project Sidewalk Labs in Toronto finally received "conditional approval" on October 31st and will finally vote in March next year after all kinds of difficulties posed by the public and the government. When completed, it will become one of the typical representatives of smart cities in the western world.

China is developing even faster. "Internet power and intelligent society" has long been a national strategy, and smart cities are an important part of it. In November, Sadie, JD.com Yun and JD.com jointly released the Research on the Development Strategy and Strategy of China's Smart City in 2019, which puts forward five strategic approaches for the development of smart city: industry, good governance, benefit to the people, symbiosis and foundation building.

The layout of domestic enterprises will also speed up in 2019.

At the beginning of the year, JD.com iCity Conference was held in Beijing, announcing that JD.com City has become a group-level strategy, intelligent Suqian APP was launched in June as the first landing achievement, and intelligent city system 2.0 was officially launched in November by JDD. Baidu is also working with Neusoft Group this year to launch a new smart city solution for "Yunzhi Future City". Aliyun partnered with Senna in May to improve Malaysia's transportation system, and in August partnered with Gaud and others to launch a joint urban brain traffic management solution.

In the future, with the introduction of more reliable devices and the development of 5G and edge computing, computing power will show a more distributed structure. The intelligent city will also develop faster, and intelligent transportation will take the lead in landing in the park, logistics, public transportation and other scenes.

05 # fake AI videos must not be released at will # see the discussion of AI regulations

In 2019, what you see is true, and it may not be true.

This year, China's ZAO App was born, which allows you to play roles in various blockbusters yourself, and has gained a large number of users as soon as it is launched. Behind this, it is the technological evolution based on deep learning that makes it possible for AI to change the faces of the characters in the video and generate all kinds of "Deepfake" videos.

While being happy with technological advances and new applications, concerns peaked in 2019.

Because this technology can also do bad things.

For example, fake speeches by politicians such as Mr Obama have even had an impact on Malaysian politics, and companies in the UK have been defrauded by lawbreakers. Regulatory and ethical discussions on AI technology have also been put on the agenda of various countries.

On November 29th, the National Information Office issued a notice that from next year, AI fake videos can not be sent at will, and then went on the hot search. The new regulations also require all major platforms to deploy "unreal audio and video identification technology" as soon as possible.

The previous month, California Governor Gavin Newsom also signed the AB-730 Act, which made it clear that it would be a crime to use Deepfake to have political influence. In July, Virginia also issued a ban on the abuse of Deepfake.

For its part, Twitter released its first draft anti-Deepfake strategy in November and solicited public comments. Giants such as Microsoft and Google have also begun to study identification technology.

In addition to the automatic generation of videos using AI, articles can also be generated. In fact, this year, the entire AI industry has made remarkable progress in natural language understanding, but fortunately, there is no problem with fake news for the time being.

In mid-February, OpenAI released a trained universal language model, GPT-2, because it was so powerful that it could generate articles that were difficult to tell true from false, and could even imitate writing habits, and its creators thought it was too dangerous to release it until November. After GPT-2 used Reddit's huge online BBS data to train and generate articles, it was highly readable and did not cause the catastrophe of fake news.

A series of models from Baidu, Carnegie Mellon University, Google and other institutions have also exceeded the NLP benchmark. In December, ERNIE, Baidu's knowledge-enhanced semantic understanding framework for sustainable learning, topped the list in GLUE, an authoritative dataset in the field of NLP, and broke through the 90 mark for the first time with an average score of nine tasks.

In addition to the fact that the final content form seen by the user may be virtual, the training data of AI may also be simulated.

Because in many scenarios, the cost of obtaining real data is too high, various organizations are exploring ways to train high-quality data models. This year, the simulation environment has been able to complete more data construction and high-quality training.

For example, the "five-man team" trained by OpenAI Five beat the world champion of Dota2, and the training of the back model was carried out in a simulated environment.

Similar simulations are more common in "dialogue" type training, such as the dialogue ability training of Amazon's Alexa, where delivery drones and robots are using simulated data.

Simulation data are also being used in the field of self-driving, with Aurora unmanned cars conducting hundreds of simulations while training models to navigate in urban environments. The Baidu team developed an augmented reality autopilot simulation system (AADS) and appeared in the Robotics sub-issue of Science magazine.

It uses lidar (LiDAR) and cameras to scan street views. According to the trajectory data obtained, the seemingly reasonable traffic flow for cars and pedestrians is generated and synthesized into the background. The resulting realistic images are fully annotated and can be used for autopilot system training and testing from perception to planning.

This year, more AI applications are also warming us.

Huang Zhisheng, who works at the Free University of Amsterdam, formed the Weibo Tree Cave rescue team to help identify and rescue people who tried to commit suicide online, which has prevented 507 suicides in the past year. Google's AI team has selected 20 public welfare and environmental protection organizations around the world with a capital injection of $25 million to help them achieve AI empowerment.

On the Baidu AI search platform, users have initiated nearly 400000 photo comparisons, and more than 9000 people who have gone missing have been reunited with their families. Baidu has also launched the world's first Mini Program sign language translation for hearing-impaired children, which can translate pictorial text into sign language and help hearing-impaired children to read without disabilities.

In the future, the ability of AI will be stronger, and with the improvement of national laws and regulations and social ethics system, AI will be safer, more controllable and more warm to human beings.

End

In 2019, topics related to artificial intelligence have been searched frequently. Such as travel, communications, and even littering, AI is closer to people's lives. However, the development of AI in China still has a significant impact on the world.

In 2019, the global AI industry has further developed to the enhancement of computing power, the industrialization of empowerment and the platform of AI. Companies such as Intel and Google, for example, enhance the computing power of AI through chips, computing centers and even quantum computing; major AI platforms such as Tensorflow continue to enhance algorithms and ease of use, which continue to widely influence academia and industry; and Baidu upgrades Baidu's brain to 5.0, passing through the whole process of AI industrial application landing, realizing the standardization, automation and modularization of AI technology, and becoming a "software and hardware integrated AI production platform".

China's AI policy advantages are gradually obvious in smart cities, intelligent infrastructure and the upgrading of primary and secondary industries, represented by Huawei, Baidu and JD.com. Both China and the United States are exploring the new possibilities of the latest AI policies and ethics.

Quantum hegemony, 5G commercial, Deepfake, these buzzwords will define 2019 of the AI situation. The prelude to a new industrial revolution with AI and other technologies as the core is under way and can be expected in the future.

Http://www.woshipm.com/ai/3277476.html

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.