(Original title: French geek Yann LeCun: At the helm of Facebook artificial intelligence)
Qian Tongxin
In the field of artificial intelligence in today's world, there are three top experts who have been regarded as “people like God†in the industry. Two of them are from Canada and one is from France. They were Geoffrey Hinton of the University of Toronto in Canada and Yoshua Bengio of the University of Montreal, and Yann LeCun (LeCun), the head of Facebook's artificial intelligence research department (FAIR). The Parisian scholar is currently a tenured professor at New York University. He is also the founder of the Data Science Center at New York University.
Yann LeCun entered the Chinese university in March of this year. He conducted two artificial intelligence top dialogues at Tsinghua University and Shanghai New York University, and accepted an exclusive interview with the First Financial reporter.
Let the machine have common sense
LeCun is a scientist who is very proud of the French academic community. He is also one of the few French people who holds a high position in the US technology giant. Although the same as "geeks", the unique temperament of the French makes LeCun more casual and full of affinity than many American scientists.
After graduating from the computer department of the sixth generation of universities in Paris in 1987, LeCun went to the University of Toronto to pursue a postdoctoral course. He studied under Geoffrey Hinton, the father of neural networks, and Hinton was also the person who brought deep learning technology to Google. After the postdoctoral research, LeCun had worked and lived in the United States and worked at Bell Labs, AT&T and other large companies. In 2008 he founded YLC, a consulting firm engaged in big data mining. Until now, he has also served as the chief scientific officer of another music-making and education company he founded.
Currently LeCun leads the team of nearly 100 people in the Facebook artificial intelligence research department. His work is to advance the basic science and technology research of artificial intelligence; to develop practical applications of artificial intelligence technology in various fields through experiments, such as computer vision, human-computer dialogue systems, virtual assistants, speech recognition and natural language processing (NLP) Wait.
"There are many basic sciences behind artificial intelligence. They may not be application-oriented. Your research may only lead to an understanding of intelligence and artificial intelligence," LeCun told a CCP correspondent.
LeCun opened up a precedent for applying neural networks to machine vision. Five years ago, it led researchers to make a huge breakthrough in the accuracy of image recognition. The technology behind this, the artificial neural network, has contributed to the prosperity of artificial intelligence in recent years and has enabled Google and Facebook to make people The search function was used in her own album and led to the launch of a number of apps that use facial recognition.
How training machines learn is the most important job for LeCun's team. For a long time in the past, they “fed†thousands of pictures to machines to teach machines to distinguish between “car†and “puppyâ€. However, LeCun also raised a new problem in the process: When there are a large number of available samples (such as tables, chairs, cats, dogs, and people), there is no problem with the training machine; but if the machine has never seen these objects, it can still identify Is there a sample?
LeCun said that a major challenge for the development of artificial intelligence is how to make the machine master human common sense, which is the key to the natural interaction between the machine and humans. To do this, it needs to have an intrinsic model to be able to predict. LeCun succinctly summarizes this artificial intelligence system with a formula: prediction + planning = reasoning. What researchers need to do now is to not rely on human training and let the machines learn to build their own internal models.
“People have spent many years researching how to automatically add captions or descriptions to pictures and videos. From the current technology point of view, there are indeed impressive implementations.†LeCun told CCP reporters, “but actually At the moment, they are not as amazing as they seem, and the professionalism of those machines is very much confined to the environment in which people train them.If you show the machine an unusual situation, most machines will be overwhelmed because They don't have common sense."
LeCun believes that there is still much room for advancement in the field of machine vision. The next breakthrough in machine vision will be to learn by watching the world autonomously, for example by watching videos. This also means that in the future computers may learn commonsense knowledge like babies learn.
Regarding how machine vision is related to common sense, LeCun said that there are even big differences within Facebook. "Some people think that they can communicate only with the intelligent system, but the language is a very low bandwidth (lowbandwidth) channel, and the information density is very low. Languages ​​can carry a lot of information because people have a lot of background knowledge, that is, Common sense to help them understand this information," explains LeCun.
Some artificial intelligence scientists believe that the only way to provide sufficient information to an artificial intelligence system is to add visual cognition, because the image is much denser than the language. For example, you tell the machine "this is a smart phone", "this is a roller", "something you can push and some cannot", etc. Perhaps the machine can learn the basic operation principle of the world. In this regard, LeCun said: "This is similar to the way infants learn. However, young children do not need explicit instructions when they are learning a lot of things." LeCun believes that learning in the absence of guidance is what he wants to achieve.
He said that what Facebook really wants to do is to allow the machine to understand the limitations of the real world by watching videos or observing other things, which will eventually make them build common sense. "Current machines are still very fools, because they lack a basic understanding of the world." LeCun said, "For example, if you look at the machine for a short video in the future, then the machine can predict what will happen next. If we can train the system To do this, then we have created the core technology of unsupervised machine learning, which is an important part of our artificial intelligence."
Convolutional neural network
Over more than 20 years of research, LeCun has published more than 180 papers. His most widely known research was the 1988 participation in the development of the famous convolutional neural network (CNN). Therefore, LeCun is also called “volume†in the industry. The father of the neural network."
Convolutional neural network is an efficient identification method developed in recent years. Its initial conception dates back to the 1960s. When researchers studied neurons used for local sensitivity and direction selection in the cat's cerebral cortex, they discovered that their unique network structure can effectively reduce the complexity of feedback neural networks. A convolutional neural network was proposed.
Nowadays, convolutional neural networks have become one of the research hotspots in many scientific fields, especially in the field of pattern classification. Because the network avoids the complex pre-processing of images, it can directly input the original images and obtain a wider range of applications. . This revolutionary system was able to recognize handwritten digits from the beginning and, as data training continues, it can begin to recognize visual features from image pixels, which is like opening up the eyes for the computer so that they can learn from the data themselves. .
LeCun told the First Financial reporter: “Now, deep-convolution networks have been used to solve various types of computer vision problems, including target recognition. And, as the depth of the network continues to increase, there are also available for image recognition, semantic segmentation, A new type of deep convolutional neural network structure such as ADAS."
Facebook is currently using machine learning to implement a range of different functions. These features include face recognition. Machines can recognize faces from the web, even if the person's face is not annotated because this technology is based on a neural network that simulates the human brain. Realized.
These networks can be trained to identify patterns in information, including language, textual data, or visual images, and are the basis for a large number of artificial intelligence research and development in recent years. The next task of the machine system will be to observe the real world and learn how the world works. One way is to learn by interacting with smart phones and wearable technologies.
In the field of artificial intelligence research for more than 20 years, LeCun's goal has always been to give the machine greater ability to make the machine smarter. He told First Financial reporters that there are many things on Facebook that they want to do and that many tasks remain to be completed. "I hope to see the application of new technologies on Facebook and make our research more meaningful. By improving the deep learning capabilities of machines, we can turn them into smart machines."
LeCun also believes that artificial intelligence can do everything in the future, including predicting people's behavior. “The next step in the machine is to be able to learn by observing everything in the real world, and to predict.†LeCun has repeatedly stressed in Twitter and Facebook that “unsupervised robots have a good prospect.â€
He believes that the biggest challenge facing Facebook when it comes to the next phase of artificial intelligence breakthrough is how to match the best content with personal needs through machine learning. Last April, Facebook launched the Chatbot chatbot at the F8 conference to help people complete tasks such as ordering and scheduling. In LeCun's view, the ultimate goal of a chat robot is to become a personal virtual assistant, to connect humans and the real world through artificial intelligence technology and to perform tasks in daily life.
LeCun told CBN reporter: “Although in the short term we can only start with some simple functional applications, our long-term goal is to build a truly intelligent machine so that you can talk directly with it and it needs to be able to answer Any problem, and help with your life.This matter is very challenging for today's artificial intelligence, man-machine dialogue system, natural language processing, all of which is based on letting the machine learn the common sense of human beings. I don't know exactly what to do, but we have many ideas for it."
Artificial intelligence is a long-term investment
In response to the current global technology giant’s fierce competition in artificial intelligence, LeCun told the First Financial reporter: “No one is running ahead. Many companies are doing a lot of artificial intelligence research and development, and competition for talent is fierce. Nobody has invented new technologies that are far ahead of other companies.†He added that no company’s new technology requires someone to spend more than three months to catch up. Everyone’s level is very close. Among the first echelons are Facebook, Google’s DeepMind, Microsoft, and IBM.
Regarding the success of DeepMind’s invention of AlphaGo, he said: “This is a great victory in the field of artificial intelligence. Some of my students and postdoctoral students have participated in the DeepMind project. This achievement is based on everyone’s efforts.†In fact, The system that analyzes the Go chessboard and determines the position of the fall is actually the convolutional neural network invented by LeCun. However, he also admits that Facebook does not have much research on Go and has a much smaller mass than DeepMind's system. “Our Go research is mainly used as a carrier for planning and exploration research. Our system works well, and then we open it up.â€
In terms of the commercialization of artificial intelligence, LeCun said: “The impact of basic research can only be realized after a relatively long time. You can't imagine planting a seed, and then suddenly emerge the physical product line, and commercial forms can be completely eliminated. Change. This is a long-term investment. It needs forward-thinking people. This kind of person has Google and Facebook also has."
Facebook recently announced that it is forming a consumer goods division. LeCun confirmed this to CBN. However, he said that the new department and the artificial intelligence department he is responsible for are two separate teams and there is no direct connection. Facebook is indeed developing artificial intelligence technologies for the consumer market. Some are software applications, some are hardware, such as AR, VR, and robots. "We are building an artificial intelligence ecosystem that can connect parts with people's lives."
LeCun advocated the opening up of research results to allow more people to understand the research they are working on. He said: “To maintain a good relationship with university laboratories, to allow these institutions to export all types of talent for you, and to conduct all kinds of possible research, we must open up projects and results. Assume that you are a scientific researcher, you are sure to always It's important for scientists to publicly publish your research results because your position is academic. You can't simply tell people 'I'm working for Facebook, but I can't tell you what I'm researching', so you His career was ruined. This is very important."
Artificial Intelligence Ambassador
The concepts of artificial intelligence such as machine learning and deep learning have gradually begun to be accepted by ordinary people, but it is still difficult for most people to really understand and express. To this end, LeCun has frequently entered global universities and colleges in recent years, and has actively promoted the science popularization work in artificial intelligence. He told First Financial reporter: “Helping the public understand artificial intelligence is very important for promoting the development of the whole industry.â€
During his trip to China, LeCun also visited the National Laboratory for Pattern Recognition at the Chinese Academy of Sciences. In a photo posted on Facebook with the researchers of the Chinese Academy of Sciences, he wrote: "It is a great pleasure to learn that China already has multiple national projects for artificial intelligence."
He said: “China’s overseas investment is a very interesting phenomenon. The Chinese company’s investment approach is basically to establish an ecological circle in the country first, and then gradually infiltrate abroad to expand overseas. In fact, when we see more and more Chinese companies invest In overseas projects, we should also see that many companies in Europe and the United States are investing in the field of artificial intelligence in China. The flow of such capital is an inevitable development of technology."
LeCun also stated that in some areas of artificial intelligence, China has surpassed the United States to lead the world. For example, in terms of deep learning, according to a report released by the U.S. government in November last year, China has published more articles than the United States.
However, compared with U.S. technology giants, there is still a gap between Chinese research and technology. LeCun believes that the artificial intelligence research laboratories in the two countries are very different. "The artificial intelligence labs of Facebook and Google DeepMind are really studying very advanced things, such as predicting the future of learning and artificial intelligence. This is what I have in other areas. No company has ever seen it."
Although the use of artificial intelligence in China has been ubiquitous, from easing urban traffic congestion to injecting transparency into the judicial system. However, the biggest problem facing China today is the shortage of professionals.
In response to fierce competition in the field of artificial intelligence, LeCun said: “China accounts for one-fifth of the world’s population and has a lot of talent here. Zuckerberg attaches great importance to the Chinese market. We also conduct artificial intelligence with Chinese universities and colleges. And there are many other basic research collaborations that are of profound significance to Facebook, but this does not mean that we have already started doing business in China."
The rapid development of artificial intelligence has brought about surprises and triggered various concerns. There is a concern that Facebook is using artificial intelligence to monitor people's behavior. In addition, with the rapid growth of artificial intelligence, many people worry that robots will soon replace humans and even control the entire world.
LeCun said: There is no need to worry so much. “Although the learning curve of artificial intelligence development security system is in the upward trend, the machine will eventually be controlled by the balance between humanities and society. Perhaps a hedge fund in a hypothetical situation can help humans to maximize profits by disrupting the economic system, but these Acts will eventually be constrained by social and legal systems."
LeCun posted on his Facebook a cartoon from the American comic book writer Bill Watterson's book Calvin and Hobbes, where six-year-old boy and tiger lay On the lawn, they cannot understand each other's world. The cartoon wrote: "If I don't know why you laugh, our life will not have much resonance." This passage is also LeCun's summary of artificial intelligence and human relationships: there are opportunities and challenges, full of charm and The color of passion, but it also gives people an unknown fear.
Fiber Optic Rotary Joint, also known as optical fiber slip ring, optical fiber rotary joint, optical hinge, etc., is used to solve the problem of optical signal transmission between opposite rotating components, that is, to ensure that the transmission of optical signals is not interrupted due to rotation.
fiber optic rotary joint,fiber optic slip ring,optical rotary joint
Dongguan Oubaibo Technology Co., Ltd. , https://www.sliproubo.com