How can we let artificial intelligence pass the period of "artificial mental retardation"?

Artificial intelligence (AI) is very hot, but many times we feel that artificial intelligence is not "smart". For example, Apple's mobile phone, the founder of a company, Yang Jing, has uninstalled Siri for a while because it is not smart enough and sometimes it will mess up. When you press it, it jumps out, making artificial intelligence seem to become "artificial mental retardation", and it feels better than not appearing.

This shows that some technical challenges have been encountered in the development and application of artificial intelligence technology. Many experts have explored AI technology in their own fields. Many explorations are very advanced. So, how can we cross this kind of way? Difficulties and obstacles, reaching a new realm of AI?

AI is so hot, the root of everything is data, huge data torrents. We are in the era of intelligent Internet of Things, everything is connected to the Internet, and more and more intelligent, are collecting data, analyzing data. Therefore, the key to the artificial intelligence technology is how to better collect and analyze the collected data, and then use it to realize value-added, which will bring value-added services. This is an important reason why people pay so much attention to artificial intelligence.

At present, artificial intelligence is still in the early stages of development. As the computer age becomes more mature, the deployment of many technologies becomes very difficult, because many technologies are under the framework of artificial intelligence, but only 7% of the entire AI-related fields It is in line with the specific requirements of AI. In order to better implement artificial intelligence, very powerful computing power is required to be able to handle such large data. It is estimated that by 2020, the computational cycle for artificial intelligence will grow by as much as 12 times.

About a year ago, Intel acquired the artificial intelligence company Saffron Technology. Saffron is a leader in platform services in the AI ​​field. They pay special attention to the processing power of technology-based logic and reasoning. They are very concerned about the analysis of data. They also hope to join everyone's understanding of data, such as the analysis of fraud and its A number of successful cases in the banking industry have been driven by Saffron Technology. Recently, the industry giant acquired Nervana Systems.

For this acquisition, Jason Waxman, vice president of Intel Corporation and general manager of the data center solutions division of the data center business unit, said that the most exciting thing about this acquisition of Nervana Systems is that we believe this company is in line with our strategy. The idea, especially in the early stages of the solution, to the entire software infrastructure, while adding to our database and helping to optimize our semiconductors. Therefore, through the acquisition of such a large number of IPs, we finally bought the entire product portfolio. After the completion of this acquisition, I feel that we have already developed this very good technology in the first 90 days. I think this is also the ability to increase Intel's processing and strategy.

At the same time, Intel also acquired Movidius, the acquisition has not yet been fully completed, but after the completion of the acquisition will be critical to the development of smart devices. Movidius is critical in the field of embedded computers, and Intel will use Movidius technology in smart camera and image recognition, as well as in the Internet of Things. Therefore, it will continue to expand its investment in this field.

Jason Waxman said that building Intel-based hardware platforms covering Xeon processors, Xeon processors, Nervana platforms and FPGAs, Omni-Path networks, 3DXPoint storage, and more, combined with Intel's optimization for deep learning/machine learning Intel Mathematical Function Library (Intel.MKL), Data Analysis Acceleration Library Intel.DAAL, etc., and open source software frameworks dedicated to providing superior performance for multi-node architectures such as Spark, Caffe, Theano, and Neon, etc. These tools and platforms, such as Saffron, TAP, Nervana system, and Movidius, which are collaboratively developed with artificial intelligence, can help enterprises more easily acquire, develop, and deploy artificial intelligence applications, and fully release artificial intelligence potential in various fields, such as Smart factories, driverless cars, sports, fraud detection, life sciences, etc.

Now that technology is evolving with each passing day, we are seeing a lot of innovations and new reforms. Some technologies have existed for several years, and now is the time to open up new market opportunities.

Machine learning has different ideas and perspectives. It usually refers to the method of learning from data. It can build this data and improve its performance through data, including how to systematically and structurally learn from the data. This is very broad. A definition of this, and this concept has been in existence for a long time in academia. Until recently, more data was generated in life and work, thus driving the need for machine learning.

There are three different types of machine learning. Naveen Rao, a former CEO and expert in artificial intelligence at Nervana Systems, said that I can't say that there are only three types, but this will give you a little insight into the world of machine learning. We can see a concept called supervised learning in many recent news. For example, I have some pictures, logos, which are posted on my data, such as a person's face recognition. The identification of this data represents his name. This name can be linked to a computer and the computer can learn. Or combine this input data with its name and picture. Unsupervised learning is at the forefront of our technology, and it's hard to really find a data architecture that is potentially usable, and only if you don't know the details of the architecture at first.

If you want to learn a language, you must first listen to it, understand its pronunciation, understand the phonetic intonation, and at least form a language framework before you can really learn the language. In fact, it is difficult for us to give machines the ability to learn, but we are still pushing this research.

Let's talk about reinforcement learning. Just like training your own pet, you want it to have a positive, desired reflection. If it gives a negative, unwanted reflection, you have to punish it. This is the concept of reinforcement learning. I think real AI is more than just these three types. In fact, making our lives and our world better is the ultimate goal of machine learning and artificial intelligence.

To give an example, it is a very traditional and classic machine learning. Here you can see some pictures for everyone. You can see that this is one of our founders. How do we teach machines to recognize people's faces, that is, let the machine connect names and faces together. According to the traditional method, I have to look at the characteristics of his face, that is, the width of the eyes and eyebrows and the length of the nose. These are very important identification features, which are identified by software and turned into an image key. Point, we call it a function of facial recognition features.

Finally, we can finally identify his name through different classifiers and some integrated methods. I think humans or animals can better distinguish this face, but machines need a lot of learning because they have no way to extract these features directly. Traditional classical machine learning is facial recognition in this way.

In the future, if you want a machine to have the same ability as a human being, what aspects does this ability reflect? Mr. Song Jiqiang, Dean of Intel China Research Institute, summarized it into four major abilities: first, the ability to perceive the external environment; second, reasoning based on the ability to perceive; third, reasoning formed some decisions to touch the machine for feedback, this feedback can be dynamic It can also be motionless, such as visual feedback and feedback from the voice; the last one is more important to adapt to the environment. It is important to adapt to the changes in the environment and the changes of the interacting person. This is very important. Otherwise it will become a dead program.

There is a horizontal line in it is very important, "memory", memory is a very prominent point. Jeff Hawkins, the founder of Numenta, wrote a book called "The Future of Artificial Intelligence," which is devoted to artificial intelligence or intelligence. He is actually able to use memory to make predictions. At this point, the machine is really smart. So the ability to remember this is very important. We now see that many manufacturers who make artificial intelligence also add this to the system solution.

The benefits of deep learning are deep, so features can be extracted in many different abstraction layers. This feature is not defined by people. Many people who used to do artificial intelligence and machine learning used to find Future, so there is Futureengineer work, but Now you don't need it, it allows features to be discovered from the data, and the performance of the system is improved by more and more data, rather than relying on one or two Futureengineers.

At the same time, deep learning is more and more powerful, from the original can only represent static images and simple sentences, now has the ability to represent continuous images, that is, video, even multi-modal . Deep learning is a breakthrough technology. Its breakthrough is that when the data is large enough, it can surpass human ability at some level. For example, in face recognition, it has surpassed people on ImageNet. ability. In terms of voice, speech recognition and machine translation have also reached a satisfactory level. Therefore, with the increase in the amount of calculation and the amount of data, deep breakthroughs have been seen in deep learning.

We expect performance to increase as the data increases. As the data increases, the training time also becomes longer. When the training time becomes longer, in one case, the performance can be extended by increasing the number of processors, so that it remains in one or two. This model can be trained in a day. If the number of processors reaches a certain level, then the bottleneck is no longer the computing power, but the I/O capability is the communication between the processor and the processor, or The I/O of the processor accessing the memory becomes a bottleneck. At this time, if you increase the number of processors, you can't see the performance increase, or it takes so much time to train. Nervana technology can more efficiently define the memory access bandwidth and computational density, making it a good match for the current data volume increase. At the same time, multiple nodes can support the training of this large model in parallel, so we can see nearly a linear performance extension.

The figure below is a tensor-based architecture. The tensor operation is a more than two-dimensional block operation, and the matrix operations are all tensor operations. These green parts are all processing units dedicated to matrix operations. At the same time, we also use a kind of Flexpoint technology. This technology is neither fixed nor double-precision floating point. It is based on the technology that can be changed between the two, so it can provide unprecedented parallelism in this. The level of computing, the calculation density is very high, is currently the best level of hardware acceleration 10 times.

At the same time, since the calculation unit is specifically designed for tensor calculations, its power consumption is very low. At the same time, we see that there are 4 yellow blocks next to it. This is a high-bandwidth memory. It is connected to the main chip through a dedicated memory access interface. The large gray area can be regarded as a chip. This memory is directly managed by the software, so there will be no Cache in the whole calculation, there will be no unpredictable Cachemiss, it is completely controlled by the program to read the data, when the data is good, I put it back.

You will think that this chip may not handle all the deep learning training tasks, then I use multiple chips. The interconnection between multiple chips is based on RCL. We have specially customized an Interchip Link, which is 20 times faster than the traditional PCIE, and is a bidirectional data bandwidth channel. With this technology, I can put a chip. It can be connected to up to 12 chips, which can form a large super mesh for training.

At the same time, with the hardware, we can't relax the research on the software AI algorithm, because the hardware provides a foundation, if you run it with a poorly designed algorithm, it is actually a waste of its resources. So in Intel Research and the software department, I also do a variety of research on AI algorithms, including these four categories. The first major category is how to make training faster, so that it requires less data and human supervision. . Because training now usually requires a lot of data, and needs good label data, but many of us who are not in the field of artificial intelligence ask, children need to identify so many pictures, give two or three pictures to understand. It doesn't take so long, so there should be more technical points to delve into. How to enhance the current deep neural network, because this is just one of the tools, not necessarily the final tool.

Deep neural network + X, what? In addition to some memory, perhaps there is a better model for learning to generate knowledge and reasoning. In addition, we know that the large model is very sparse after training. There are many parameters in it. It is worthless to transfer it or store it even for these 0s, so we have to find a way to make the model more sparse. Let it be stored more compressed, and we can do the cropping in the model. When cutting, some 0 places will not be operated, so the amount of calculation will be reduced.

Another dimension is the decrease in the accuracy of the calculation. This is also a hot research point. We used 16 bits to calculate. Now let's see if we can use 8-bit, 4-bit, 2-bit or even 1-bit. The calculation of the model, at the same time, I will ensure that the performance and accuracy will not drop too much, which is also a focus of our research. At the same time, I can see how it can be extended, and use the larger Bach Size and High-order methods to do deep learning training.

This year, starting with AlphaGo, AI has become a social issue. Now that AI is very hot, its application is still limited to some narrow areas of expertise, such as image recognition, speech recognition, or driverless. Intelligent driving, neurolinguistics, etc. So, is there any related application of AI in public life?

For the application of AI in people's lives, Zhang Wei, chief engineer and artificial intelligence solution architect of Intel Data Center Division, said that the application of deep learning is still in the infant stage, and we expect it to be in various fields in the next few years. There will be more in-depth and broader applications. Intel is currently mainly engaged in image, voice or natural language processing, especially with our cloud providers (Cloud Provider).

In the next step, we will go deeper into the application of the industry, such as the application in the medical field. This is closely related to the life of all of us. For example, the identification of medical images, many times the machine judges cancer, I have seen domestic The professors have done a very good job in this area. They use machine learning to judge cancer, which is more accurate than the average doctor.

Shangtang Technology is mainly engaged in artificial intelligence landing products, such as the field of live broadcast. In this regard, Liu Wenzhi, the engineering director of the company's heterogeneous parallel computing department, said that it is very important for female anchors to be beautiful. Shangtang's face recognition technology can beautibly beautified. The value of the female anchor, which is actually used in a wide range, of course, also borrowed the live broadcast of this year.

In addition, we are used in some face gates for customs clearance. If we pull out our ID card or other equipment ourselves, it is relatively complicated. If we rely on our own face, it can be unimpeded, including Under certain circumstances, it is not necessary to bring a key to your home. There are also some image related, of course, Shang Tang Technology can also provide a good solution. For example, Pikazo.

Chen Yu, director of research and development of artificial intelligence in Jingdong Group's AI/VR/AR laboratory, said that as a technology R&D practitioner, I feel that this matter is divided into two layers. First of all, from the perspective of perception, perception itself is a very important area of ​​artificial intelligence. The direction, such as vision, hearing, our analysis of things, that is, image recognition, speech recognition, we are all landing.

For example, our image recognition is not just some face recognition or text recognition that we have seen, or some simple identification of images. We can even help our warehouse system to review and sort goods. It is a specific landing scene. In addition to perception, we also have many other applications in Jingdong. For example, Jingdong has Jimmy robot, which can replace Q&A services manually, which can save a lot of labor. These are not just perceptions, but also intelligent information. Processing, these are the actual landing scenes of artificial intelligence in Jingdong, and things that everyone can directly understand.

Dr. Feng Xiaobing, a researcher at the Institute of Computing Technology of the Chinese Academy of Sciences and a doctoral tutor, deputy director of the State Key Laboratory of Computer Architecture, said that because I am doing research, I can only say some examples of others in specific applications, like us. Now using smart phones, handwriting recognition, voice input, this is followed by artificial intelligence support, including like Jingdong, many domestic e-commerce websites, product recommendation, and behind it are all with the help of artificial intelligence, let It is done more accurately. Including some travel, simple question and answer system, customer service, etc., I believe that there are artificial intelligence support behind.

But I think that in the future, we should have a role in the new manufacturing and industrial upgrading.

Feng Xiaobing said, I think there are two major trends in the development of artificial intelligence. The first is how to make more applications. From the theory and method of artificial intelligence itself, it can handle more applications. Now it is mainly based on the learning method. In addition to the learning method, we have nothing else to let it go, but this may not have much to do with us. We may be similar to Intel, because the computing system is all system, so the second big problem is how do we make a more reasonable system, so that the existing artificial intelligence application makes it acceptable. a wider range of applications under the premise of the cost,

Support for chips, system software, including programming frameworks, and many of these challenges may face many challenges. Regardless of whether the calculation is done or the business community mentioned, it is trying to make a better system, so that we can use smaller power consumption, lower program development cost, and get better performance. Do a larger data size process. On the other hand, because there are many different kinds, do we have some basic technologies that allow some systems to get better adaptability to different application types without requiring too much intervention from programmers or system engineers? This will face great challenges in the future.

In 2020, where will artificial intelligence appear to be newer and more compelling breakthroughs?

In this regard, Feng Xiaobing said that my most hope is still in the medical field, because medical care does have a lot of news events, which is very important for every family. Through artificial intelligence, it can be used as a medical auxiliary and management system. There may be no way to completely replace the doctor, but we can use the assistance of artificial intelligence to make our average doctor's standard and the level of medical services that everyone receives better, which will make everyone feel more at ease and spend less money. Getting better service is my expectation.