"Small data" is gradually becoming popular, behind which is the natural result of technological evolution or the re-selection of "big data" after hitting a wall? What is "dark matter" in the cognitively intelligent world? How should we expect and think about ai-general?
"We were the first team in the world to study big data problems, and we started doing research in the field of big data almost 17 years ago (2004). After about three or four years of research, it was found that there were some inherent problems in big data, which were foreseen to be unsolvable by perceptual intelligence. Later, we began to experiment with paradigm shifts, starting in 2009 to study cognitive intelligence. Recently, Dong Le, executive vice president of the Beijing General Artificial Intelligence Research Institute, said in an interview with the surging news (www.thepaper.cn).
Beijing General Artificial Intelligence Research Institute is positioned as a non-profit new R & D institution, supported by the Beijing Municipal Government and the Ministry of Science and Technology, supported by Peking University, Tsinghua University and other units, and prepared by Professor Zhu Songchun, a world-renowned computer vision expert, statistical and applied mathematician, and artificial intelligence expert, in 2020 and served as the dean. The goal is to achieve universal agents with autonomous perception, cognition, decision-making, learning, execution, and social collaboration, consistent with human emotional, ethical, and moral concepts.
Dong Le elaborated that at present, we see that more AI adopts the "big data + computing power + deep learning" paradigm, which belongs to the intelligence of the perception layer, and when the real industry lands, the current perceptual intelligence has encountered many problems, such as being able to do tasks defined in advance by specific humans, there is a long tail effect, high training costs, a large number of data labeling involves privacy and security issues, in addition to the unexplainable, inexplainable, non-communication, algorithm bias and other problems of the model.
"Now everyone has slowly formed a consensus that cognitive intelligence may be the direction of artificial intelligence development in the next 10 years." Professor Dong Le said.
How to understand cognitive intelligence and perceptual intelligence?
"Crow Paradigm" and "Parrot Paradigm"
A Japanese wildlife scientist collected many videos of the daily habits of wild crows. He found that when a wild crow came to the city, it needed nuts to be full but had no way to open them. At this time, it had a very accidental discovery, throwing nuts on the road, and after the car drove past, the nuts were crushed and could be eaten directly.
But it faces a new problem in the process of eating, the road is very dangerous, how does it complete this task? Very cleverly, it found the signal light again, when the red light, all the cars stopped, it threw the nuts on the zebra crossing, the nuts were crushed by the wheels, and when the signal lights indicated, the cars stopped and then came down to eat the nuts.
"All of this series of actions is done autonomously, by solving a task — eating nuts safely, it observes, reasons, discovers the laws of traffic, and then executes and makes decisions." We call this the 'crow paradigm', the "small data, big task" paradigm. It doesn't have a high training cost and doesn't require much data training, but it has to accomplish a mission goal, so it's task-driven. Dong Le said.
The opposite of the "crow paradigm" is the "parrot paradigm", and the parrot needs a lot of data to train repeatedly, teaching him what it says. It can be repeated continuously, but it does not understand its meaning, it cannot reflect the causal logic in reality, and it is a "big data, small task" paradigm.
Under the perspective of cognitive intelligence, the three key elements of ai systems are "architecture, tasks, and data." Dong Le believes that compared with the emphasis on "data, computing power and models" of perceptual intelligence, this is another step forward. Among them, architecture is the most important. "Just like the ability to judge a person, not from how much knowledge he has mastered, but from his complete knowledge to build models, then even if the current knowledge is not enough, but to a new field, with such a sound architecture can also quickly acquire new knowledge." We believe that architecture is the foundation and mission is key, and data plays part, but not all, of the role in that process. ”
For example, train the AI to complete the task of chair recognition. If you follow the paradigm of perceptual intelligent deep learning, you need to label the features in a large number of chair images, and then let the AI learn. However, after that, when encountering a special-shaped chair, it is still difficult to identify. "Not only in simple object recognition, but also in areas including unmanned driving and medical treatment, similar problems will be encountered." Dong Le said.
But people don't need to see a lot of chairs and it is easy to make a judgment about whether it is a chair or not, and how do people do it?
Dong Le summed it up, "We people will raise this task from a simple object recognition problem to a high level of understanding of the task. Judging through visual perception and physical imagination, that is, when we see it, we can imagine whether it can withstand making me sit safely, and sitting on it is uncomfortable, it's as simple as that. ”
Dong Le once mentioned the "dark matter" in the cognitive intelligent world in the forum of the BEYOND International Science and Technology Innovation Expo. She believes that in daily life, it is easy for us to perceive the input of information from the senses such as vision, but this is only the tip of the iceberg. "The reasoning and imagination behind the senses actually exert a huge amount of energy, which we call 'intelligent dark matter'. We will understand and reason about physical and social common sense, and then combine space-time and causal models to act in real scenes, and integrate perception and cognition. ”
AI can learn the abstract ability of human beings to extract invisible knowledge, based on the paradigm transformation of "Dark Beyond Deep", that is, to complete the "big task" through a small amount of data, with a small number of samples, simple annotation, to do a hundred examples, to understand the world in a way that combines perceptual intelligence and cognitive intelligence, and to explore intelligent "dark matter".
For "small data" to become increasingly popular, behind which is the natural result of technological evolution? Or is it a re-choice after "big data" hits a wall? Dong Le believes that there are two levels.
"We don't deny big data, big data does have great value in many scenarios, but what to do in other scenarios? At the same time, there are data problems, cost problems, energy consumption problems... Using big data to solve some problems that can be solved without big data is actually very unscientific. Dong Le told the surging news (www.thepaper.cn).
If you roughly compare the effectiveness of the parrot paradigm crow paradigm, Dong Le said, "The parrot paradigm may be 2:8, that is, the universal ability is only about 20%, and it also needs to be personalized according to the task requirements; the crow paradigm is 8:2, the universal ability reaches 80%, and only 20% of the ability needs to be optimized and iterated according to the task requirements." ”
As for whether to recognize the research route of brain-like intelligence in the path forward of artificial intelligence, Dong Le told the surging news (www.thepaper.cn), "If you put aside the problems and tasks to be solved and simply discuss a technical paradigm or a path, I think it is of little significance and value." Every technical path has a certain necessity for its exploration and research, simply to say which path may have problems, or some people have questions, this is not surprising, the key is to solve what problems, to determine the task. ”
Dong Le used the analogy of mountaineering, there are many roads from the foot of the mountain to the top of the mountain, the surrounding scenery is also different, the problems to be solved in the process are different, and now when looking up from the bottom of the mountain, there is no way to judge which road is the best. Maybe only after you have really reached the top, you can go back and think about this problem.
Is general artificial intelligence "artificial intelligence" like people?
In a 2014 interview with the BBC, physicist Stephen William Hawking expressed concern about a "human-like" AI, "Making machines that can think is undoubtedly a huge threat to the existence of human beings themselves." When artificial intelligence is fully developed, it will be the end of mankind. ”
In the years since, Hawking has also expressed this view in several speeches. In 2017, Hawking warned in an interview with the British newspaper The Times that "further development of artificial intelligence may destroy humanity through nuclear or biological warfare." Humans need to use logic and reason to control future threats. ”
So at the moment, when we discuss general artificial intelligence, we are pointing to the artificial intelligence that Hawking is worried about?
Zhang Cymbal, academician of the Chinese Academy of Sciences and dean of the Institute of Artificial Intelligence of Tsinghua University, once said at the Fifth Chinese Intelligent Conference, "The development of general artificial intelligence is a good thing, and it is also a happy thing that it has really developed, but general artificial intelligence and strong artificial intelligence cannot be confused here." ”
Zhou Zhihua, dean of the School of Artificial Intelligence of Nanjing University, once described "strong artificial intelligence" as an artificial object that reaches or even exceeds the level of human intelligence, an "artificial intelligence" with mind and consciousness, and can act according to its own intentions in the "Column" of the China Computer Society Newsletter in the first issue of 2018. "General artificial intelligence" hopes to learn from human intelligent behavior and develop better tools to reduce human intellectual labor, its essence is behavioral intelligence and task intelligence, and the essence is still "weak artificial intelligence", similar to "advanced bionics".
"The progress and success of artificial intelligence technology is due to the research of 'weak artificial intelligence' rather than 'strong artificial intelligence'," Zhou Zhihua said, "from a technical point of view, the efforts of the mainstream artificial intelligence community have never been towards strong artificial intelligence, and the development of existing technologies will not automatically make strong artificial intelligence possible." ”
Michael Wooldrige, former president of the International Federation of Artificial Intelligence and head of the Department of Computer Science at Oxford University, said in a report at the 2016 CCF-GAIR conference that strong AI has "little progress" and "little serious activity."
"General AI is task-driven and currently within limited boundaries, and just like us humans, human capabilities have boundaries." Dong Le told The Paper (www.thepaepr.cn).
What kind of general artificial intelligence can be achieved? Dong Le believes that it is actually a mission, a direction, which constantly allows agents to solve problems in a more general way. The first manifestation is that the agent can have the ability of common sense reasoning in the general sense, and about 80 or 90% of the tasks can be accurately understood and realized. The second is that a technique is basically universal in scenarios with the same logic.
"For example, in the fields of medical care, education, finance, including energy, there are a large number of resource matching problems, decision makers need to make real-time predictions based on limited information, so the analysis is not fast, accurate, accurate and comprehensive, to analyze the causes, so that it can be clearer and more reasonable," Dong Le said, "Our cognitive AI universal agent, its role is actually to give these comprehensive information to the hands of those who need it more rationally, help decision-makers, better and fairer overall planning, allocation of resources." , make the most scientific decisions. ”
At present, many enterprises are also using artificial intelligence to complete intelligent transformation. During the reporter's visit, he found that many companies in transition are hesitant about whether to build their own AI teams. "At present, we will see that many national enterprises are also facing such problems, one is whether the data can be given, and the second is whether their professional ability can be enough." Dong Le told The Paper (www.thepaper.cn).
Dong Le believes, "If it is just an enterprise application, you should cooperate with a professional team." AI talent itself is very scarce, the cost is very high, if there is no strong scientific research and engineering capabilities, you will eventually find that the input is more and more, but the output is not obvious. If you think about laying out your own AI team from the perspective of enterprise strategy, this is another problem. Purely from the output orientation of the results, I think that for most enterprises, there is no need to form their own professional AI team, find an excellent professional team, build a good cooperation model, and do what they are good at is the optimal solution. ”
Talking about the help of AI to human and social welfare, Dong Le said that in fact, it is to use technology to break the unbalanced, wasteful and depleted resource allocation methods, so that the overall operation efficiency of society can be intelligently improved. "We believe that in the next 50 years, there will be a collision and integration of artificial intelligence and human civilization. In fact, for all social administrators, including each of us, we must think about what we have to face in the society of the intelligent era. ”