(Original title: Shen Xiangyang, head of Microsoft AI: current level of AI, technical difficulties, the road to Microsoft AI, and how to face the talent challenge)
On May 10, 2017, just after the keynote speech at the Microsoft Build 2017 conference, Dr. Shen Xiangyang, head of the Microsoft Artificial Intelligence and Microsoft Research Division, was interviewed by InfoQ and other invited media. This article organizes the main points of this interview.
Why is artificial intelligence so hot now? Shen Xiangyang said: “The powerful power of cloud computing and the powerful algorithms that run on deep neural networks, coupled with the massive amounts of data we can obtain today, are driven by the interweaving of these three powerful forces. Today, we are finally able to realize Smart dream. Artificial intelligence has infinite potential, it has the ability to subvert any existing vertical industry."
So, what is the current level of artificial intelligence development? Where are the technical difficulties on the way forward? How is Microsoft's artificial intelligence approach planned and how does it face the competition for talent?
First, how to look at artificial intelligence?
In answering questions from InfoQ reporters, Dr. Shen Xiangyang fully explained his understanding of artificial intelligence: artificial intelligence is divided into two aspects of perception and cognition, and there has been significant progress in perception, and the cognitive aspect is still far from breakthroughs. The interpretable AI will achieve a major breakthrough in 5 to 10 years.
1. Why artificial intelligence?
When it comes to artificial intelligence, although everyone will be thrilled today - I also think - some things are already happening, but first of all you have to look back, why artificial intelligence?
Artificial intelligence comes from human intelligence, which is Human Intelligence. Later, 60 years ago, John McCarthy defined a word called Artificial Intelligence, and it is said that McCarthy really provided such intelligence.
2. Artificial intelligence is divided into two parts: perception and cognition
Why do people think that people have intelligence? In fact, human intelligence is basically divided into two parts, one is perception and one is cognition, and artificial intelligence is also corresponding.
(1) Great progress in perception
The greatest and greatest part of perception is visual perception.
Someone had done such research before. One person's 91% of information was collected from the visual. I forgot what method he used to calculate this number, but everyone basically agrees: Most of the perceptions come from For vision, then hearing, and finally for other perceptions.
I think the progress here is very large. I have always said that in the past year I have been talking about computer speech recognition, that is, five years of things. Within five years, computer speech can be recognized. No matter how you say it, you can Identification. In the next 10 years or so, I think that computer vision will reach this point. Today, many things in vision have surpassed people and face recognition. What I'm talking about is a general sense of universality. To a new place, to see what a new thing can think of, these things can take about a decade or so to come to fruition.
(2) The cognitive aspect is far from breakthrough
In terms of cognition, we are far from making breakthroughs today and we are not talking about human beings.
The first is the problem of natural language processing, and then the problem of knowledge acquisition. More and more people should do more work in this area. Natural language processing, I have just mentioned that machine reading and language are relatively complicated. With today's existing methods, including deep learning methods, the results obtained are still not good enough. Of course, using deep learning can already help us with many things. For example, translation also uses many natural language things.
More importantly, today's definition of the entire "cognitive" thing is still at a relatively elementary stage.
For example, what is Common Sense? How do you know that after seeing this person, why would he be very close to him? We do not understand this yet, and this is a big problem.
Slightly aside, one of the most important questions is that today's personal intelligence is not enough to combine these things with brain science and understanding is not enough. The main reason is that there are only a few examples of the "intelligence," such as the human brain. However, the human brain has a very special structure. Today we don't understand enough. As a science, brain science is still in an early stage today. We can't do too many experiments, and we can't open a person's head and stuff stuff into it at any time.
This is a long-term problem, and now more and more people are thinking of such issues - connecting artificial intelligence and brain science.
(3) Cognitive, interpretable AI will make great achievements
There is one direction that makes us feel very excited: Now many people at the Microsoft Research Institute are doing this, and I also have some cooperation with many universities, the so-called Explainable AI. In my opinion, Explainable AI can certainly make great achievements in the next 5 to 10 years. If I have a graduate student today, I will let them do the work in this direction. The reason is very simple, because the biggest breakthrough of AI today is deep learning, but one of the biggest problems of deep learning is that the results are very good, but you Can not explain.
The best written article in this area, which I have seen myself, is a recent article in the New Yorker about medical AI. Why do you see the same picture, the doctor will tell you, you have no problem, the reason is one, two, three. But AI still cannot do this today. Deep learning can't do this. The big problem is that the space for everyone to solve the problem is not the same. The doctor is in such a continuous space as the so-called neural, brain. A lot of AI's understanding is done in the discrete space of symbols.
Therefore, how to link these things up and technically speaking, there are many areas that need to be broken through, and this is also one aspect of our current research institute that is very serious in scientific research.
Second, Microsoft's progress in artificial intelligence
1. Voice aspects
First of all, we talk about the breakthrough of artificial intelligence in speech. Artificial intelligence has recently achieved very noticeable results in speech recognition and speech synthesis. In September 2016, Microsoft's dialogue speech recognition technology achieved a word error rate as low as 6.3% in the industry standard Switchboard speech recognition benchmark test, creating the lowest error rate in the field at that time. A month later, Microsoft further reduced the word error rate to 5.9%, for the first time reaching a level with the professional stenographer and outperforming the vast majority of people.
2. Image aspect
Second, in terms of images, artificial intelligence also has a lot of progress:
In December 2015, the ImageNet computer vision recognition challenge was unveiled. The researchers of the Visual Computing Group at Microsoft Asia Research Institute relied on the latest breakthroughs in deep neural network technology to gain absolute advantages in image classification, image positioning, and image detection. The champion of the project. At the same time, they also succeeded in another image recognition challenge MS COCO (Microsoft Common Objects in Context). In the ImageNet Challenge, the research team at Microsoft Asia Research Institute used an unprecedented neural network with a depth of 152 layers, which is more than five times the number of successfully used neural network layers in the past, so as to identify photos and video objects. Other technical breakthroughs have been achieved, reducing the error rate to 3.57%.
In October 2016, researchers from the Visual Computing Group at Microsoft Research Asia won first place in the Image Recognition MS COCO Image Segmentation Challenge, achieving 11% more performance than the second, and compared to the COCO image segmentation challenge of the previous year. The first place in the game has also seen dramatic progress.
3. Natural language aspects
In addition to voice and images, artificial intelligence has made great progress in natural language.
(1) Man-machine dialogue: In terms of human-machine dialogue in natural language, deep neural networks have gradually replaced traditional statistical machine learning and become the mainstream research direction. Nowadays, natural language technology has all turned to deep learning networks, and our dialogue systems have also used deep learning networks. The well-known Microsoft Xiao Bing, the key natural language processing technology that is used by Microsoft's natural language processing technology, has now been able to realize 23 conversations with humans.
(2) With respect to machine translation, Microsoft Translator now supports more than 60 languages ​​and can realize multi-language real-time translation for multiple people. For example, everyone can come from different countries. As long as they hold the Microsoft Translator for mobile phone APP, they can interact with each other. communicate with. When you say a word or enter text, what the other party hears/sees is his mother tongue. At the just-concluded Microsoft Developer Forum 2017, Microsoft also showed the latest Presentation Translator PowerPoint plug-in, which leverages Microsoft's Translation API interface to translate real-time presentations into presentations. multilingual.
(3) In terms of machine reading comprehension, at the SQuad (Stanford Question Answering Dataset) Text Understanding Challenge initiated by Stanford University's Natural Language Computing Group, Microsoft Asia Research Institute's Natural Language Computing Research Group continued to top the list. The Microsoft Research Asia team achieved the best results in the two different dimensions of accuracy and similarity. The accuracy reached 76.922%, and the similarity reached 84.006%, which was higher than the second place. Percentage.
Third, Microsoft's artificial intelligence research
Any company or unit, especially after it is big, must think about what the short-term goal is and what the long-term vision is. We must think from this perspective. My department is special because, besides AI, I also manage the AI ​​Institute. At the AI ​​Institute, we have more than 1,000 scientists and we must continue to train generations of new and amazing researchers to do even more remarkable technology.
1. Microsoft's four major directions in the field of artificial intelligence
First, the search engine. The world’s largest artificial intelligence may still be a search engine today. Microsoft Bing has surpassed 2.5 billion entities for so many years. There is a lot of knowledge here, search engine itself is not just a business, although now Bing is also very profitable, we in the United States 22.6% of the search market share plus Yahoo's 11% (technology is our back office to do), so We have a 1/3 search share in the United States and rose to 16.5% in China recently. From the perspective of AI, it is the accumulation of knowledge.
The second, very, very important thing is Cortana. I think Cortana represents the future of AI and the understanding of people. To do a good job of AI requires three aspects of knowledge:
Understanding of the world
Understanding of work
Understanding of users
If these three pieces are added together, they can do very well. I think Xiao Na is moving in this direction. Of course, there is a lot of investment in this task.
Third, other companies and Microsoft cooperate together, how to make AI help Microsoft transition, I just mentioned Office, also mentioned Cloud, also mentioned Windows, we do together. At the same time, we put some of these technologies out to all other Microsoft developers. The end point that I talked about today has been made in our AI department for so many years. There are many technologies that have been done from Microsoft Research Institute for several decades. .
Fourth, the mining of business opportunities. All commercial applications will be overturned. In which direction Microsoft is choosing, what business AI opportunities we will tap, and we hope to have the opportunity to share the progress with you.
2. Long-term training of talents for scientific research
Any company or unit, especially after it is big, must think about what the short-term goal is, what the mid-term hope is, what the long-term vision is, and we must think from this perspective. What makes my department special is that I manage the Institute in addition to AI. At the Institute we have more than 1,000 scientists. Just now Xin Zhiyuan's colleagues also asked, I think the most important thing is to continue to train new generations of amazing researchers and do even more remarkable techniques.
For example, I just mentioned the integration of artificial intelligence (more symbolic processing in discrete space) and brain science (more continuous processing in Neural neurons). The relationship between the symbolic interpretable space and the continuous brain space, basically no one studied these directions three or five years ago. I think the more important thing is to cultivate a new generation and see such problems. If you study deeply, you have to explain it.
Fourth, Microsoft's road to artificial intelligence products
1. Why must have a product?
For the general public, he will not read the paper to judge your research results, nor does he know how great your paper is. You want to explain to people, the easiest way to explain is that you give people a look at your product, HoloLens will be immediately seen, we hope to have the opportunity to do something even more remarkable, not only to consumers, but also more. Product for business users. Personally, I think that within 3-5 years, the biggest opportunity for AI is still in the corporate market.
2. Correctly look at scientific research
Many people do not understand how to say that so many people do scientific research. Scientific research is a very long-term thing. You should either say that you do scientific research or you need to be patient with scientific research. For example, today we talk about the sky-rocketing quantum calculations, which specific days of quantum computers can be made, no one knows. At this point, the United States system still deserves our study: from the university, to the institute, to the industrial sector. Former AT&T and IBM, and now Microsoft, many companies are willing to spend a lot of money to do long-term research, and the vast majority of scientific research results do not only belong to the company, but their own company may benefit.
In this regard, Gates made it very clear. For example, Apple and Microsoft's early success was very important to the graphical user interface. Graphic interface was first made by Xerox. We learned from them. Similarly, Microsoft has done a lot of amazing things today, but it may be normal for other companies and even some startup companies to do better.
Moreover, doing research is a very pleasant thing. The happiest thing to do in scientific research is to not worry about what others are thinking. You can pat your brain and think about it. There is an amazing idea of ​​how much I used to be. He enjoys doing scientific research and was eventually driven out of Ballmer to make products.
3. How to go from technical research to productization
Of course, we are not a public welfare research institute. We are responsible for the company, including the most important thing, that is, the transformation from technology to products. At present, AI itself is still studying a lot of things today, so today Microsoft reorganized and put AI and the research institute in the same department - I feel very honored to lead such a department - we see so many opportunities.
How to transfer from technology to product, perhaps the best example today is cognitive services, of which about two-thirds of the technology was originally done by Microsoft Research, and it has been done for many years. We were not very clear before. Some computer vision How the technology is transformed into a product. However, because of Azure, there are opportunities for cognitive services. Many of the technologies of Microsoft Research have been transformed into products through cognitive services.
Another example is HoloLens, whose R&D process is a process of “research, development, research, and developmentâ€. HoloLens and these people used to be Kinect. Kinect made Microsoft Research Institute to do Kinect Fusion, and also made a project called Holodesk. If it is three-dimensional, how do you add some three-dimensional virtual objects? Afterwards, Microsoft had a number of great engineers who worked on products and design. They thought that HoloLens should be built on it. During this process, many of the computer vision and voice technologies were developed together by Microsoft Research. The process of research and development.
4.Product planning of AI department
It is very important to earn some money and set a small goal. However, the more important thing is that after the establishment of the AI ​​department, we have to think clearly if we really think that AI will subvert more industry applications. In the process of subversion, where are our opportunities?
(1) AIization of existing products
For example, when Office and AI are combined, what disruptive content will come out, what new products will come out, what new features will come out, and we will make very good progress here.
The three main aspects of AI are: First, you must have very strong computing power; Second, you must have a very good algorithm; Third, you must have your own data.
I use Microsoft's example to introduce it. At Microsoft, we certainly believe that all of Microsoft's products must be AI-enabled and define such products again.
We are now focusing on two areas. One is on all Office products. In the keynote, I showed everyone a function of PowerPoint, which is translation. In fact, they also did another PowerPoint. I liked it very much. The so-called image caption: To come up with a photo, PowerPoint shows that the production will automatically give you a picture, which we have done quite well.
PowerPoint is used by many people. This data can help us continuously improve some algorithms. Not long ago, we also released Word, which uses AI technology in Word. This is also very important.
There are still many AI technologies here that have only just begun. I think that the most exciting technology for Office is the so-called machine reading. Not long ago, Microsoft bought a Canadian start-up called Maluuba, mainly to do this work, using natural language, deep learning to do this thing. One of the most important questions in deep learning is to answer questions. I think the impact on Office will be huge, so our colleagues in the AI ​​department are working with colleagues in Office.
The other one is in the cloud. We have a lot of cooperation. You go to the Azure.com homepage and the content of Cognitive Services is placed in the most prominent position. This is Scott and I, after reviewing with our product team colleagues, Cognitive Services will become the top priority of Azure. Windows also has a lot of AI, like HoloLens also has a lot of AI technology, computer vision, computer voice.
(2) Excavating new product lines: deciding what to do and not doing
The other one is to think about where the new product line is, whether you have a new product line, and you can do billions of dollars in business after three or five years. To think about it, if you have such a business, five to ten years can be a business of 10 billion dollars. If there are, of course, we must put the horse to pursue such an opportunity.
So the most important thing for us now in the entire AI department is to decide what to do and decide not to do anything.
Fifth, how to deal with artificial intelligence talent challenges?
1. How to look at the flow of talent?
I think that the flow of talent is very normal. A large company has trained many talents. The most important thing is to say that any company you want to know the value and philosophy of your existence shows that why good employees choose to stay with you, not only When you go outside to dig people, what kind of amazing environment you provide to the staff here, so that he has a very good development here.
When I last reported with Secretary Liu Yunshan in Wuzhen, I told everyone that Microsoft in China has trained countless talents for the IT industry in China. I said that you should not only see the CTOs of almost all IT companies in China train me. , from Lenovo to Haier to a small company, I said that you have to see that Microsoft Research has trained 5,000 students in the past 18 years. Those are really amazing. A new generation of startups is coming out - it may not be possible now. This is the situation - when computer vision and AI company first got up, those investors came to ask me some situations. I said you don't want to tell me that the computer vision company in China is opened by my students. Either it's my student's student, it's a very glorious thing. Looking back, Microsoft Research and Microsoft have played a huge positive role in China's IT development in many aspects, and especially trained a lot of first-rate talent for China.
Microsoft Research has always felt very proud about exporting talent. We trained Li Kaifu. We trained Zhang Yaqin. We trained Zhao Feng. We trained Yong Yong, Yong Yong, or my brother. All of these were very good. We all feel very honored.
2. How to train talent
One of the most important things we have done recently is about talent. Just six months ago, shortly after the establishment of our department, I established the Microsoft Artificial Intelligence Academy and trained a considerable number of Microsoft internal talents. We hope that in this way we can attract more and more outside talents to come to Microsoft. In doing so, it is not that we are worried that some people will dig our AI talents. What's more important is to develop our talents into AI capable. Therefore, we have a prefixed lesson, a double prefix lesson, and a three-headed lesson. To the six-character class, I recently completed a course on AI611, specifically for deep learning. There were 10 projects. It was amazing. At the end of the course, I spent 2 hours listening to their reports. Very good! Therefore, we are currently training AI talents, both internal talent training and attracting outside talents.
6. Conclusion
A few years ago, it was hard to imagine that the same technical tools would be driven by artificial intelligence.
After a few years, it is hard to imagine that there will be no artificial intelligence behind any technology.
With the growing power of cloud computing, powerful algorithms that run on deep neural networks, and the massive amounts of data that can be acquired today, driven by the interweaving of these three powerful forces, today we are finally able to realize the dream of artificial intelligence.
Artificial intelligence has endless potential, it has the ability to subvert any existing vertical industry, such as banking or retail, and any single business process, such as sales, marketing or human resources and headhunting.
In this way, one day, artificial intelligence will have the ability to add brilliance to man’s boundless intellect—enhancing human capabilities and help us achieve greater productivity.
Respondent profile:
Shen Xiangyang, Ph.D., Executive Vice President, Microsoft, Head of Microsoft Artificial Intelligence and Microsoft Research, Academician of the American Institute of Electrical and Electronic Engineers, Academician of the American Computer Association. He is responsible for Microsoft's global artificial intelligence strategy and is responsible for forward-looking research and development covering infrastructure, services, applications, and smart assistants. He is also responsible for the group of artificial intelligence products, including the Microsoft Information Platforms Division, Bing and Xiaona Product Division, and the Environmental Computing and Robotics team. In addition, Dr. Shen Xiangyang is also responsible for the integration with Microsoft's product engineering department.
燑br>
We specialize in cable assembly overmolding. Our customized molded cables are built precisly following customers' spec.Experienced proposal can be offered to customers for evaluation by considering stable quality and competitive price.
Custom cable overmoldings improve the quality, look and functionality of many electrical cables. From battery interconnects and molded plugs to custom molded plugs and many other custom designed assemblies – ETOP is a premier custom cable overmolding supplier to OEMs.
we have internal design, prototype, and manufacturing the widest range of molded harnesses, also a diversified line of strain / flex reliefs and grommets.Molded Cables,Stranded Round Molded Cables,Molded Patch Cord Cable,4 Pole Waterproof Molded Cable,Waterproof Connectors
ETOP WIREHARNESS LIMITED , https://www.oemmoldedcables.com