Preview in the future: Microsoft Asian Institute will look at a twenty year

 The process of becoming a networker isn’t considered for the faint-hearted. It requires lots of hard work and nice and trustworthy DevNet 300 915 DEVIOT Dumps, like that offered at the ITCertDumps, to clear this grueling exam.

(above is Microsoft Global Senior Vice President, Madam President of Microsoft Asia Research Group, Mini Shanzhi, Ministry of Asian Research, Hong Xiaowen)

2018 On November 8, Microsoft Asian Institute has a 20th anniversary celebration. In November 1998, the Microsoft Asian Institute was established in Beijing, which is Microsoft's largest research institutions outside the United States. For 20 years, Microsoft Asia Institute has developed into computer foundation and application research institutions with world influence. At present, Microsoft Asian Research Institute has more than 200 researchers, and more than 300 visits and interns, mainly focused on natural user interface, intelligent multimedia, big data and knowledge excavation, artificial intelligence, cloud and edge computing, computer science foundation Six major research areas.

As of November 2018, the Microsoft Asian Institute has published more than 5,000 papers in the international top academic conference and journals, including more than 50 "Best Paper" awards, recently announced artificial There are 27 of the Microsoft Asian Research Institute. In addition, many global scientific and technological breakthroughs also from the Microsoft Asian Research Institute: 2015 Microsoft Asian Research Institute developed computer vision transcendence to the ability of human objects in the ImageNet Challenge; 2018 in Stanford University Squad On the text understanding challenge list, the R-NET and NL-NET models of the Microsoft Asian Institute take the lead in the two dimensions, and March 2018, the Microsoft Asian Research Institute and the Microsoft Research Institute The machine translation system jointly developed, in the general news report test set newstest2017 - the first time I have been comparable to artificial translation.

When it is constantly breaking through the scientific research border, the Microsoft Asian Institute is also deeply influenced by Microsoft's product system. From Microsoft Intelligent Cloud Azure, Office 365 to Microsoft, Bing Search, then to Xbox and Hololens, you can say that Microsoft has a Microsoft Asian Institute of Asia, Microsoft Asia Research Institute also passes itself. The technical accumulation and research innovation has hatched a lot of popular applications and technology platforms. From the talented talents from the Microsoft Asian Institute, almost supported the innovation army of the China Science and Technology Industry, including Li Kai Fu, Zhang Yongqin, Wang Jian, Ma Weiying, Qiang Yong, Li Shipeng, Wang Haifeng, etc. is the top in ICT, Internet and investment. Level leadership science and technology talents.

Microsoft Asian Research Institute has also hatched a number of engineering institutes, including Microsoft China Cloud Computing and Artificial Intelligence Division and Microsoft (Asia) Internet Engineering Institute, which not only promotes the development and landing of Microsoft products, but also further improve Microsoft's R & D layout in China and Asia Pacific promoted the establishment of Microsoft Asia-Pacific R & D Group. In September 2018, Microsoft Asian Institute announced in Shanghai to establish Microsoft Asia Research Institutes - Shanghai, and announced the establishment of Microsoft-instrumental artificial intelligent innovation institutions with the Shanghai Xuhui District People's Government and Shanghai ViDe (Group) Co., Ltd..

In the past two decades, Microsoft Asian Research Institute has made excellent achievements. In the next 20 years, how do Microsoft Asia Institutes? On the 20th anniversary of the establishment of the Microsoft Asian Institute, the Microsoft Asian Institute of Asia has more than ten artificial intelligences such as machine learning, computer vision, system research, data intelligence, personalized recommendation system, natural language processing, computer graphics. The core technical direction issued an article and interprected the views on the technology trends in the next decade to twenty years. This article is selected, which has been published six, and sharing the "foresee of the future" of the Microsoft Asian Institute of Asia.

Good system "Elephant invisible"

In the entire computing research and product innovation system, computer systems are the most basic, and most important areas. Computer system research involves very broad, including software-level operating systems, database systems, compilation system lines, including hardware levels of CPU instructions, memory, storage systems, servers, and data center structures.

Microsoft Asian Research Institute believes that "large sounds, elephant invisible" is a "good system". "Intangible" means that modern computing system design is naturally accompanying the public, but it also allows the public to "turn it", which is like unwritten gas stations or power interfaces, which block complex urban oil supply. And power supply lines.

In the visible future, the systematic research will still use "invisible" as an important design concept, allowing the application developer and the public without the face of the complexity of the underlying system. The future technology trend is a process of disappearing in real world and virtual world borders, or it can be simply "virtual and real-world fusion". These changes require a large number of individual types of can be embedded in a connected sensor and execution device, intelligent data analysis and services, and support for mixing reality and immersion experience equipment.

In the cloud computing system architecture, an obvious technical trend is how to evolve from complete centralized clouds into new detrimentary calculations. For example, the rise of IoT and edge computing is generated in this context. More importantly, from the computer architecture and philosophical category, the deposition has always been a system design that dominates the whole world. Of course, it is also unrealistic to completely go to centering. In the future, the data and calculations will seamlessly migrate, interactively migrate, interactive, and collaboration will become an important direction of system research, from operating systems, storage systems, and even consistency agreements, further improvements.

Future calculation is getting higher and higher in real-time, how to target design corresponding computing platforms, is a topic of system research. The development of storage platforms, from BigTable, MongoDB to Spanner, follows to improve scalability as the primary demand, but the requirements for consistency are increasing, high-scalable and strong storage platforms will be the future trend. one. How to optimize the performance and energy consumption of the system in intelligent edge computing is a topic worth continuous research. In addition, in the edge calculation, the model is stored and running on the edge device. Therefore, how to protect the model data is not illegally replicated and pirated, it has become an important issue. In terms of security, privacy, and credible calculations, security hardware, block chains, and verifiable software technology will form a new security basis.

The integration of software and hardware is a future important trend. How to better define the boundaries between hardware and software and the appropriate abstraction, how to find the best division between them has become an important issue for system architecture design. For example, for the introduction of deep learning, Tensor and the calculation map, how to cooperate with software and hardware to complete this new computing model; the abstract combination of data operations, data operations will also be further improved Cloud computing platform handles the efficiency of big data. In the new integrated design revolution, new specific field programming languages, compilers and optimization, and virtualization technology will also change at the same time.

The heterogeneous system is a hot spot in computer architecture research in recent years. Depth learning Accelerator, FPGA-based reconfigurable hardware, universal programmable accelerator, and new general purpose processors are constantly innovating. The innovation of memory hardware, such as high-speed nonvolatile memory (NON VOLATILE MEMORY, NVM), memory package innovation, such as HBM (High Bandwidth Memory), and architectural innovation, such as memory decomposition, will also help calculate from memory Access the bandwidth and delayed bottlenecks are liberated.

The front end expressive ability of future artificial intelligence computing systems will become more flexible, gradually tend to make common calculations; at the same time, the calculation capacity of the backend will become more and more powerful. In the face of the rapid development of the back, the automated compilation optimization framework has become the passive path of the connection. In the artificial intelligence era, large-scale computing systems must not only efficiently handle massive data in large-scale equipment, but also support the capabilities of multiple different types of tasks. In the future, different computing tasks will become increasingly blurred in the actual boundary, and the data analysis in many real-world applications is a complex consisters of these computing tasks. Now more large systems have more randomness, control variables and environment variables have increased significantly. These new changes allow artificial intelligence to play a greater role. In the future, system researchers need to explore how to apply a more wide-rate-based approach to a large complex system design and control.

What are the machine learning?

In recent years, there have been many new types of machine learning technologies to be widely concerned in people, but also provide effective programs, including deep learning, strengthen learning, confrontation, dualism, migration Learning, distributed learning, and meta. Although machine learning has achieved a lot of practical issues, there is still a huge challenge in the field of machine learning.

First, mainstream machine learning technology is a black box technology and cannot predict the hidden crisis. To solve this problem, you need to make the machine learn interpretable and intervene. Second, the current computational cost of mainstream machine learning is high, and it is urgent to invent lightweight machine learning algorithm. In addition, in physical, chemical, biological, social sciences, people often use some simple and beautiful equations (such as second-order partial differential equation such as Schr?dinger Corcation) to describe the deep law behind the appearance. So in the field of machine learning, is it possible to pursue simple and beautiful laws? There are still many such challenges, but the Microsoft Asian Institute is still confident in the future development of this field.

Among them, in terms of lightweight machine learning and edge calculation, edge calculation refers to processing, analyzing data, and analyzing data, and the edge node refers to data generating source and cloud computing center. There are nodes with computing resources and network resources. For example, the mobile phone is the edge node between the human and cloud computing center, and the gateway is the edge node between the smart home and cloud computing center. In the ideal environment, edge calculation refers to analysis, processing data, reduces data, and reduces network traffic and response time. With the rise of the Internet of Things and the wide application of artificial intelligence in mobile scenes, the combination of machine learning and edge calculations is particularly important.

It is worth mentioning that quantum machine learning, this is a quantum calculation and the cross discipline of machine learning. Quantum computers use quantum coherence and quantum entanglement to handle information, which has the essence of the classic computer. The current quantum algorithm has exceeded the best classic algorithm in several issues, we call quantum acceleration. When the quantum calculation encounters machine learning, it can be a mutually beneficial reciprocity, complementary process: one aspect can utilize the advantages of quantum calculations to improve the performance of the classic machine learning algorithm, such as efficient implementation of the machine learning algorithm on the classic computer on the quantum computer. On the other hand, the quantum computing system can also be analyzed and improved using a machine learning algorithm on a classic computer.

, in terms of simple and beautiful, facing a complicated phenomenon and system, the former has got an unexpected conclusion: it seems that complex natural phenomena is made of simple and beautiful mathematics. The law is engaged, such as partial differential equations. Mathematica's creator, well-known computer scientists, physicist Stephen Wolfram have also given similar observations and conclusions: "In fact, almost all traditional mathematical models in physics and other science are ultimately based on partial micro-assembly equation." Since nature The law behind the phenomenon is so common, so can you design a method to move the law of learning and discovering the mathematical law behind the phenomenon? This problem is obviously difficult, but it is not completely impossible.

Impromptive learning is another interesting research direction of machine learning. Predict learning using all currently available information, based on past and now predicts the future, or is based on the current analysis. The problem is whether the world of life can predict? The answer to this question is unclear. Different from predictive learning to the world, improper learning assumptions are normal. Impromptive intelligence refers to the ability to resolve problems when they encounter unexpected events. Improved learning means that there is no definitive, preset, static optimized goal. Intuitively, the impromptuity of uninterrupted, self-driven ability is improved, rather than the optimized gradient generated by the preset target. In other words, improper learning through autonomous observation and interaction, to achieve knowledge and solve the problem.

Social machine learning makes the machine to simulate the behavior of human society. Currently, the purpose of machine learning is to simulate human learning processes. But so far, machine learning ignores an important factor, which is human social attribute. Since human intelligence is inseparable from society, can you make the machine also have a sense of social attribute, simulating key elements in human society, and achieve more effective, intelligence than the current machine learning method, can be explained "Social machine learning"? In fact, the current machine learning method has begun to appear "social intelligence" sporadic shadow. Since social attributes are human nature attributes, social machine learning will also use machine learning from access to artificial intelligence to acquire social intelligence.

Circular path to AI happiness

Artificial intelligence is changing our lives, more changing or subverting all walks of life. Artificial intelligence also has a lot of challenges to the human world while creating opportunities.

Challenge 1: Machine or replace human beings, resulting in unevenness and rich and poor. Challenge 2: Artificial intelligence threats human own safety and social order. Challenge 3: Artificial intelligence or promoting information monopoly or opinion two poles. In the face of the challenges of these artificial intelligence, the problem has changed from the initial "what it can do" in order to "what it should do". Artificial intelligence has created many opportunities and has also brought uncertainty. The open attitude should be kept, and the spirit of questioning is also retained, facing these challenges and seizes the opportunities.

Dealing with the AI ??challenge, Microsoft Asian Research Institute proposes to consider the following path:

path 1: Promoting calculation of thinking popularization education, let everyone become AI's understanding and beneficiaries. Behind the concerns of computer intelligence beyond human intelligence, presenting the lack of thinking culture in basic education. Let the new generation of educated people learn the basic concepts of computer science to solve problems, design systems and understand human behaviors, and become the key to establishing its future competitiveness.

Microsoft's goal is to make artificial intelligence. Since 2014, Microsoft Asian Research Institute actively cooperates with the Ministry of Education and Big primary and secondary schools to calculate thinking as the entry point, and learn from the education of education, reform the computer basic education model. In the past five years, nearly 140 related courses have been opened in 110 universities in 29 provinces and cities across the country, and the key universities in the eastern United States have driven local universities in the western teaching resources, and millions of teachers and students. At the same time, through "Innovation Cup", "Programming Beauty", "Programming One Hours" and other series of competition activities, let the concept of "calculating thinking" are higher education, the basic education community pays great attention to its promotion. The background of informationization and intelligence puts forward new requirements for talents, the importance of computer basic education and calculating thinking is also increasing, and Microsoft is creating more possibilities for this new era with scholars at home and abroad.

path 2: Promote the digital transformation of various industries, so that each company has become AI users and creators. Technical changes will be wrapped in industry development. We see the extensive application of artificial intelligence in the field of education, medical care, pension, environmental protection, urban operation, judicial services, etc. The core of artificial intelligence is data, and each industry has a large amount of data, and if the artificial intelligence technology effectively uses these data, industry innovation will not be just a vision.

The connotation of digitized transformation is far more than the traditional business digitization, which is only the first step in the journey, and intelligent is the "distant" we hope to arrive. Artificial intelligence should be deeply involved in all walks of life, combined with specific application scenarios, thus thoroughly transform the traditional industry working model and product form. For example, Microsoft Asian Institutes and Huaxia Fund have developed strategic cooperation research on artificial intelligence in the field of jackets, and jointly developed English learning and application of "Lang Wen Xiaoying" with the east, and the Oriental Overseas Shipping (OOCL) Through the application of artificial intelligence research, the shipping network operation is improved to improve efficiency. In the medical field, Microsoft Asian Institutes work together with Pfizer to construct coronary heart disease, stroke, hypertension, high blood fat, etc., intellectual maps and intelligent question and answer systems, to help patients communicate more effectively with doctors.

path 3: Strengthen national and government guidance, allowing AI more effective under policies, laws and norms. In many challenges in artificial intelligence, technology is not the final problem, and people's initiative is more worthy of attention. After several generations of people, there are countless outstanding scientists, and must be better management and norms to truly benefit mankind. We also need a force to alleviate the impact, policies and laws of science and technology to human society.

In the three major challenges mentioned earlier, many problems can also need to help solve the country, such as unemployment, unbreaking, security, monopoly and prejudice, education, etc. need to be powered. In terms of unemployment, according to the experience of the previous industrial revolution, the intelligence of the industrial revolution, the acceptance of people's new things will increase. But unfortunately, human progress is not synchronized with scientific and technological progress, which has a time difference between this. How to make this time a smooth transition, let people who are impact scientific and technological can have other choices, which requires the attention and help of the country and the government. At the same time, how to effectively eliminate data barriers, enhance the user experience, which also requires government efforts.

path 4: Urges technology companies to self-discipline, let Ai are more secure and reliable, more transparent, and more explanatory. Technological companies are the beneficiaries of artificial intelligence technology, should also shoulder the social responsibility of technology applications, so that artificial intelligence has become more reliable. Since the study of artificial intelligence, Microsoft has always focused on the moral ethical ethical issues of artificial intelligence development and application. As a joint founder, the Artificial Intelligence Cooperation Organization (PAI) is committed to promoting the discussion of relevant issues.

Personalized recommendation system

According to reports, the recommendation system brings 35% of sales revenue to Amazon, bringing up to 75% of consumption to Netflix, and Youtube homepage 60 % Browse from the recommended service. How to build a valid recommended system, the meaning is far-reaching. Microsoft Asian Research Institute from the application of deep learning, the application of knowledge maps, the application, user portraits, and the recommendation of the recommended system, and look at the future of the recommended system.

Depth learning technology is very broad in the recommended system. One of the challenges is diversified data fusion: the data of users or items is often complex and diverse, and the contents of the items may include data such as text, images, categories, and the user's behavior data can come from social networks, search engines, news reading applications, etc. Multiple fields, users' behavioral feedback can also be used in search, browsing, clicking, collection, purchase and other behaviors in e-commerce websites. In these different dimensions, data distribution of different users or items is different; the amount of data on different behavior feedback is different, and the amount of data for clicking behavior is often much greater than the amount of data purchased. Therefore, a single and group model is unable to effectively handle these diverse data, how to deliberate these complex data is a technical difficulty.

In terms of capturing short-term preferences for users, users can be divided into two categories for long-term and short-term, long-term preferences often refer to users' long-term interests, short-term preferences referring to users in the current environment Interest, such as popular songs that prefer in the last week. How to combine the impact of contextual factors, combine users' long-term preferences, effectively, and effectively combined, and also the research hotspots of recommendations.

Compared to social networks, knowledge map is a heterogeneous network, so it is more complicated and sophisticated to design a recommended algorithm for knowledge maps. Introducing knowledge maps into the recommended system, mainly as two different processing methods: feature-based knowledge map assistance recommendation, the core is the introduction of knowledge map characteristics; structurally-based recommended models, the structural characteristics of knowledge maps are used. For example, the Microsoft Asian Institute uses an outward transmission method, and historical interests of each user as the seed collection on the knowledge map, it is an extension extends along the link in the knowledge map.

By fusion depth learning and knowledge map technology, the performance of the recommended system has been greatly improved. However, most recommended systems are still established in one step. Here, it is often necessary to assume that user data is fully acquired, and its behavior remains stable over a longer period of time. However, for many real-world scenarios, such as e-commerce or online news platforms, users and recommended systems often have a continuous interaction behavior. In this process, user feedback will compensate for possible data deletions, and strongly reveal their current behavioral characteristics, providing a more accurate personalized recommendation for the system to provide an important basis.

One of the core tasks of building a recommendation system is how to accurately analyze the characteristics of the user's interest, that is, the user portrait. Simply put, user portraits refer to tabs from various data generated by users, such as age, gender, occupation, income, interest, and so on. At present, the mainstream user portrait method is generally based on machine learning, especially those with supervising learning, but still face challenges, especially how to build depth, unified and dynamic user portraits from multi-source integrated user data, which can be expandable The research direction includes: a user representation model with more characteristic ability; user portrait based on multi-source and heterogeneous data; sharing and user privacy protection of different platform user portrait data; the unified user of the user's portrait represents the model.

The research mentioned above, mostly focusing on improving the recommendation accuracy, but is not enough to communicate with the recommended object. Can I fully grasp the user's psychology in a manner that is easy to accept, and give appropriate examples to communicate with the user. The study found that such systems can not only improve system transparency, but also improve users' trust and acceptance of the system, the user chooses the probability of recommending products, and the user satisfaction, designing such an interpretable recommended system is the ultimate goal. At present, Microsoft Asian Research Institutes have considered three aspects of the following three aspects: using knowledge map enhancement algorithm interpretation capabilities; model irrelevant can interpret the recommended framework; combined with the generation model for dialogue, such as working with Microsoft Xiaobi, Generate music recommended interpretation, etc.

Image recognition: Opportunity and Challenges

In computer visual field, image recognition has developed out of the year. The high-value applications of image identification techniques are around, such as video surveillance, automatic driving, and intelligent medical care, and these image recognition of the latest progress is deep learning. Although it has achieved great success in image recognition in image recognition, there are still many challenges that need us to face before it is further widely used. At the same time, it also saw a lot of research directions with future value.

Challenge 1: How to improve the generalization of the model. Image Identification Technology is prior to wide application, how can an important challenge can know that a model is still good for unabled scenes.

Challenge 2: How to use small size and oversized data. While deep learning has achieved huge success by using a large number of labeling data in various tasks, existing technologies often crash in small data scenarios because there are only a few markings available. This scenario is often referred to as "few-shot learning" and needs to be carefully considered in practical applications. Another extreme is how to effectively improve the performance of the identification algorithm using large-scale data, and the current algorithm does not use this oversized data.

Challenge 3: Comprehensive scenario understanding. In addition to identifying and positioning objects in the scene, humans can also infer the relationship between objects and objects, partially to the overall level, and the properties of the object and the three-dimensional scene bureau. A wider understanding of the scene will help, for example, the application of robots interact, because these applications typically require information other than the object identification and location. This task not only involves the perception of the scene, but also needs to understand the real world. To achieve this goal, there is still a long way to go.

Challenge 4: Automation Network Design. In recent years, image recognition of the center of gravity of this field stepping from the design and updated network architecture. However, the design network architecture is a lengthy process that requires a large number of super-parameters and design choices. Tuning these elements requires experienced engineers spend a lot of time and effort. More importantly, the optimal architecture of a task and another task may be completely different.

Data Intelligent Future Hot Point

Data Intelligent Research Fitting in the fields of today's large data age, various industries are excavated, realize value, and the urgent need for digital transformation, so in recent years It has been fully emphasized and develops quickly. Outlook the future, data intelligence technology will continue to develop in more automatic, smart, more reliable, more general, more efficient.

Hotspots 1: Analysis at a higher semantic understanding level. In order to more intelligently analyze the data, there is a need for a richer semantic understanding of data. Unlike knowledge maps, although the most common relational data model in data analysis is also modeling entities and relationships, the modeling of relational data models is optimized for query and storage performance, often lost a lot of semantics. information. How to introduce domain knowledge and common sense knowledge, which is critical to better understanding data.

Hotspots 2: Construct the framework of universal knowledge and model. Humanity can give an antipyrettical three, touch the category. Specifically to the field of data analysis, the knowledge and models used in the analysis need to share and migrate between different data objects and analysis tasks. In the field of machine learning, there have been many related work, and some methods are proposed, such as migration learning, multi-task learning, pre-learning models, and more. To achieve this goal of this "one to three", in addition to the need to study the specific machine learning algorithm, it is also necessary to think from model and knowledge framework system, study the original knowledge and model of the field of data analysis, and knowledge and models Migration sharing unified framework.

Hotspots 3: Establish high quality training data sets and benchmark data sets. Due to the lack of training data, artificial intelligence, deep learning and other technologies have encountered great difficulties in the data intelligence. As IMAGENET data has played a significant driving effect on the research of computer visual fields. The research of data intelligence is also urgent to establish a set of public large-scale, high-quality training data sets and benchmark test data sets. Once there is a rich training data, there are many research in data intelligence, such as automatic analysis, natural language interaction, visual recommendation, etc., will make breakthrough progress.

Hotspot 4: Provides an interpretable analysis result. Users will no longer satisfy only black box type intelligence, end-to-end acting on the entire task, and require finer granularity, targeted, more transparent data intelligence. For example, data intelligence is used in a financial audit system to accurately recommend the most risky transaction rec