Is "small data" becoming more and more popular behind the natural result of technological evolution or the re selection after "big data" hit the wall? What is "dark matter" in the world of cognitive intelligence? How should we expect and think about general artificial intelligence?
"We are the first team in the world to study big data, almost 17 years ago (2004) Began to do research in the field of big data. After about three or four years of research, it was found that there were some inherent problems in big data. At that time, it was predicted that these problems could not be solved with perceptual intelligence. Later, we began to try paradigm shift and began to study cognitive intelligence in 2009. " Recently, Dong Le, executive vice president of Beijing general artificial intelligence research institute, said in an interview with surging news.
Beijing general artificial intelligence research institute is positioned as a new non-profit R & D institution. It is jointly supported by the Beijing municipal government and the Ministry of science and technology, supported by the cooperation of Peking University, Tsinghua University and other units. Professor Zhu Songchun, a world-famous computer vision expert, statistical and applied mathematician and artificial intelligence expert, prepared and served as the president in 2020. Its goal is to realize a general agent with independent perception, cognition, decision-making, learning, execution and social cooperation ability, which is in line with human emotion, ethics and moral concepts.
Dong Le explained that at present, we can see that many AI adopt the paradigm of "big data + computing power + deep learning", which belongs to the intelligence of perception layer. When the real industry is launched, the current perception intelligence has encountered many problems, such as only doing specific tasks defined in advance by human beings, long tail effect and high training cost, A large number of data annotation involves privacy and security problems. In addition, there are various problems such as unexplainable model, non communication, algorithm bias and so on.
"Now we have gradually formed a consensus that cognitive intelligence may be the key development direction of artificial intelligence in the next 10 years." Professor Dong Le said.
How to understand cognitive intelligence and perceptual intelligence?
"crow paradigm" and "parrot paradigm"
A wild zoologist in Japan has collected many videos of the daily habits of wild crows. He found that when wild crows came to the city, they needed nuts to eat, but there was no way to open them. At this time, it had a very accidental discovery. It threw the nuts on the road. 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 clever, it found the signal light again. When the light was red, all the cars stopped, and it threw the nuts into the zebra crossing. The nuts were crushed by the wheels. When the signal light indicated, the cars stopped and ate the nuts.
"All this series of actions are completed by itself. By solving a task - eating nuts safely, it observes and infers, finds the law of traffic, and then performs and makes decisions. We call this the 'crow paradigm', that is," small data, The "big task" paradigm. It doesn't have high training cost and doesn't need too much data training, but it needs to complete a task goal, so it is task driven. " Dong Le said.
The opposite of the "crow paradigm" is the "parrot paradigm". Parrots need a lot of data and repeated training. They can say whatever they are taught. It can be repeated continuously, but it does not understand the meaning. It can not reflect the causal logic in reality. It is the paradigm of "big data, small tasks".
From the perspective of cognitive intelligence, the three key elements of artificial intelligence system are "architecture, task and data". Dong Le believes that compared with the "data, computing power and model" emphasized by perceptual intelligence, this is another step forward. Among them, architecture is the most important. "Just like judging a person's ability does not come from how much knowledge he has, but from his ability to build a complete knowledge model. Even if he doesn't have enough knowledge at present, he can quickly acquire new knowledge in a new field with such a sound architecture. We believe that architecture is the foundation, task is the key, and data plays a part in this process Role, but not all. "
For example, training AI to complete the task of chair recognition. If we follow the paradigm of deep learning of perceptual intelligence, we need to mark out the features in a large number of chair images, and then let AI learn. However, after that, it will still be difficult to identify when encountering special-shaped chairs. "Similar problems will be encountered not only in simple object recognition, but also in areas such as unmanned driving and medical treatment." Dong Le said.
But people don't need to see many chairs, and it's easy to judge whether it's a chair. How do people do it?
Dong Le summarized, "we will raise this task from a simple object recognition problem to a high level of understanding of the task. We can judge it through visual perception and physical imagination, that is, when we see it, we can imagine whether it can bear to let me sit safely and feel comfortable. It's that simple."
Dong Le once mentioned "dark matter" in the world of cognitive intelligence in the forum of beyond international scientific and technological innovation Expo. She believes that in daily life, we can easily perceive the input of sensory information such as vision, but this is only the tip of the iceberg. "The reasoning and imagination behind the senses actually play a huge role. We call it 'intelligent dark matter'. We will understand and reason about physical and social common sense, and then combine the space-time and causal models to integrate perception and cognition in the real scene."
AI can learn from the abstract ability of human beings to extract invisible knowledge. Based on the paradigm transformation of "dark beyond deep", that is, complete "big tasks" through a small amount of data, use a small number of samples and simple labels to make an example, understand the world by combining perceptual intelligence and cognitive intelligence, and explore intelligent "dark matter".
What is the natural result of technological evolution behind the growing popularity of "small data"? Or the re selection after the "big data" hit the 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 about 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 surging news.
If we roughly compare the effectiveness of parrot paradigm and crow paradigm, Dong Le said, "parrot paradigm may be 2:8, that is, the general ability is only about 20%, and 80% of the ability needs to be customized according to the task requirements; crow paradigm is 8:2, the general 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 way of artificial intelligence, Dong Le told surging news, "If we put aside the problems and tasks to be solved and simply discuss a technology paradigm or a path, I think it has little significance and value. Each technology path has some necessity of exploration and research. It is not surprising to simply say which path may have problems or some people have doubts. The key is to solve what problems and determine the task.
”
Dong Le compares climbing to climbing. There are many roads from the foot of the mountain to the top of the mountain, the surrounding scenery is also different, and the problems to be solved in the process are also different. Now when looking up from the bottom of the mountain, there is no way to judge which road is the best. Maybe only when you really get to the top, you can turn back and think about this problem.
is general artificial intelligence "artificial intelligence" like human beings?
In 2014, physicist Stephen William Hawking expressed his concern about a "human like" artificial intelligence in an interview with the BBC. "Manufacturing machines that can think is undoubtedly a huge threat to human existence. When artificial intelligence develops completely, it will be the end of mankind."
In the following years, Hawking also expressed this view in many speeches. In 2017, Hawking warned in an interview with the times that "the further development of artificial intelligence may destroy mankind through nuclear war or biological war. Mankind needs to use logic and rationality to control possible threats in the future."
So now, when we discuss general AI, do we point to the AI that Hawking is worried about?
Zhang cymbal, academician of the Chinese Academy of Sciences and President of the Artificial Intelligence Research Institute of Tsinghua University, once said at the Fifth China artificial intelligence conference that "the development of general artificial intelligence is a good thing, and it is also a happy thing if it is really developed, but we can't confuse general artificial intelligence with strong artificial intelligence here."
Zhou Zhihua, Dean of the school of artificial intelligence of Nanjing University, once described "strong artificial intelligence" as an artificial intelligence that reaches or even exceeds the level of human intelligence, has mind and consciousness and can act according to their own intentions in the column of the first issue of communication of China Computer Society in 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 its essence is "weak artificial intelligence", which is 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. "Technically, the efforts of the mainstream artificial intelligence academic community have never been towards strong artificial intelligence, and the development of existing technology will not automatically make strong artificial intelligence possible."
Michael wooldrige, former president of the International Federation of artificial intelligence and head of the computer department of Oxford University, once said in the report of the 2016 ccf-gair conference that strong artificial intelligence "has made little progress" and "little serious activity".
"General artificial intelligence is based on task driven. At present, it is carried out within limited boundaries. Just like us, human capabilities also have boundaries." Dong le on surging news (www.thepaepr. CN.) express.
What can we do to achieve what we really call general artificial intelligence? Dong Le believes that it is actually a mission and a direction, which constantly enables agents to solve problems in a more general way. The first embodiment is that agents can have common sense reasoning ability in a general sense, and about 80-90% of tasks can be accurately understood and realized. The second is that a technology can be generally used in scenarios with the same logic.
"For example, there are a lot of resource matching problems in the fields of health care, education, finance, including energy. Decision makers need to make real-time predictions based on limited information. Therefore, the analysis should be accurate, accurate and comprehensive in the future. It is necessary to analyze the reasons, so as to be clearer and more reasonable," Dong Le said, "The role of our cognitive AI general agent is actually to give these comprehensive information to people in need more reasonably, help decision makers, better and more fairly plan and allocate resources, and make the most scientific decisions."
At present, many enterprises are also completing intelligent transformation with the help of artificial intelligence. During the interview, the reporter found that many enterprises in transition are hesitant about whether to build their own AI team. "At present, we will see that enterprises in many countries are also facing such problems. One is whether the data can be given, and the other is whether their professional ability can be." Dong Le told surging news.
According to Dong Le, "If it's just an enterprise application, you should cooperate with professional teams. AI talent itself is in short supply and the cost is very high. If you don't have strong scientific research and engineering ability, you will find that the input is increasing but the output is not obvious. If you want to layout your own AI team from the perspective of enterprise strategy, this is another problem. Simply from the output orientation of results, I don't think it is necessary for most enterprises to set up their own professional AI team. Finding an excellent professional team, building a good cooperation mode and doing what they are good at is the best solution. "
When it comes to AI's help to human and social welfare, Dong Le said that in fact, it is to use technology to break the unbalanced allocation of resources that may lead to waste and loss, so as to intelligently improve the overall operation efficiency of society. "We believe that in the next 50 years, there will be a collision and integration between artificial intelligence and human civilization. In fact, all social administrators, including each of us, should think about what we are going to face in the society of the intelligent age?"
