Recently, in an interview with reporters, Professor Zhang Bo, academician of the Chinese Academy of Sciences and dean of the Institute of Artificial Intelligence of Tsinghua University, believes that the current technology of artificial intelligence based on deep learning has reached the ceiling. From a long-term perspective, we must follow the path of human intelligence, and ultimately we must develop a world where humans and machines coexist harmoniously. In the future, it is necessary to establish interpretable and robust artificial intelligence theories and methods, and develop safe, reliable and credible artificial intelligence technology.
Simple Application Server
USD1.00 New User Coupon
* Only 3,000 coupons available.
* Each new user can only get one coupon(except users from distributors).
* The coupon is valid for 30 days from the date of receipt.
Academician Zhang Bo: AI miracle is difficult to reproduce in the short term, and the potential of deep learning technology is close to the ceiling
Three years after Alphago and South Korean Go player Lee Sedol won, some signs gradually appeared. Academician Zhang Bo believed that it was an appropriate time and accepted the interview.
Deep learning is currently the most concerned area of artificial intelligence, but it is not the whole of artificial intelligence research. Zhang Bo believes that although there is still room at the industrial level, the current deep learning-based artificial intelligence technology has reached the ceiling. The "miracle" brought by this technical route has not appeared after Alphago's victory, and it is estimated that the future will be very good. Difficulties continue to appear in large numbers. It is difficult for technological improvement to completely solve the fundamental flaws of artificial intelligence at the current stage, and these flaws determine that the space for its application is limited to a specific field-most of which are concentrated in image recognition and speech recognition.
At the same time, in Zhang Bo’s view, the current global business community and some academic circles are too optimistic about the judgment of deep learning technology. Artificial intelligence urgently needs to be promoted to a new stage, and this is destined to be a long process. Combining mathematics and brain science to achieve breakthroughs in underlying theories.
As a rare researcher in China who has experienced two stages of artificial intelligence technology, Zhang Bo has rarely accepted interviews in the past few years. One of the reasons is that he has some different views on the current state of artificial intelligence technology development. Before the time came, Zhang Bo cautiously believed that these views were not convenient for dissemination through mass media, and even if they were disseminated, it was difficult to gain recognition.
1. "The miracle did not happen, and according to my estimation, it will not continue to happen in large numbers"
Economic Observer: How do you estimate and evaluate the current status of artificial intelligence development?
Zhang Bo: This round of artificial intelligence boom began at the beginning of this century. The first is to appear in academia. Academia used to be cold-hearted about artificial intelligence, but the emergence of multi-layer neural networks has brought some changes. The theory of neural networks existed in the 1950s, but it has always been in a shallow application state. People did not expect multi-layers. What new changes will it bring.
What really caught everyone’s attention was the 2012 Stanford experiment (Note: In 2012, Google and Stanford used a multi-layer neural network and a large amount of data for image recognition experiments). In the past, the number of image samples in the experiment was at most "10,000". We used a multi-layer neural network to do 10 million, and found that in the three image categories of human face, human body, and cat face, the recognition rate of this model is about 7%-10% improvement.
This gives everyone a very big shock, because usually the recognition rate needs to be increased by 1%. It takes a lot of effort. Now only the number of layers has been increased. Two major changes have occurred. One is that the recognition rate has increased so much; the second is that it can handle this. Big Data. These two changes are very encouraging to everyone, let alone before 2012, artificial intelligence has not solved practical problems.
Economic Observer: What is the reason for this breakthrough?
Zhang Bo: Now we have analyzed three reasons. Everyone is very clear. One is big data, one is computing power, and the other is algorithm. After realizing it, the industry and outsiders were shocked by deep learning overnight, and then three historical events occurred.
The first thing is that in December 2015, Microsoft reduced the image recognition error rate to 3.57% through a 152-layer deep network, which was lower than the human error rate of 5.1%; the second thing was the speech recognition that Microsoft did in 2016 , Its word error rate is 5.9%, which is the same level as a professional stenographer; the third thing: Alphago defeated Korean Go player Lee Sedol.
Through artificial intelligence, deep learning and big data are two tools that can surpass humans under certain conditions and in certain fields. These three things give everyone a great encouragement.
Especially for people outside the industry, they think that as long as I master big data and use deep learning, I might be able to do miracles. So everyone has made many predictions, such as how short a computer will be able to do. More than people.
But in fact, after this, the miracle did not happen, and according to my estimation, it will not happen in large numbers in the future. To be precise, progress may be made in individual areas in the future, but it will not fully blossom as previously expected. In particular, the Chinese market is optimistic that "the Chinese market is large, data is abundant, and its use is not restricted, so miracles will definitely happen in China in the future."
As a result, many companies found that it was not the case when they were doing it. Judging from the current situation, the best results are still two things: image recognition and speech recognition. I took a look. Of the 20 unicorns and 30 quasi-unicorn companies in the field of artificial intelligence in China, nearly 80% are related to image recognition or speech recognition.
Economic Observer: Why is there such a situation? In other words, after so long, do we have a clear understanding of what artificial intelligence can currently do?
Zhang Bao: After artificial intelligence defeated humans on Go, this panic occurred. "What a master can do, artificial intelligence can actually do. My job is so ordinary, it will definitely be replaced by a machine." Here you need to consider its limitations. I have been talking at various conferences about not being too optimistic.
The three things that artificial intelligence can do (speech recognition, image recognition, Go) is because it meets five conditions, that is, as long as these five conditions are met, the computer can do well, as long as there are any one or more If the conditions are not met, the computer will be difficult to do.
The first is that you must have sufficient data. Sufficient is not only a large number, but also diversity, not incomplete, etc.
The second is certainty.
The third one is the most important. It requires complete information. Go is a game of complete information. Cards are a game of incomplete information. Although Go is complicated, it only needs to be fast in nature, and don’t rely on intelligence, but in daily life. , All our decisions are made under incomplete information.
The fourth is static, including evolving according to the law of determinism, which is the problem of predictability. Autonomous driving under complex road conditions does not meet this requirement; in fact, it does not satisfy certainty or complete information.
The fifth is a specific field, if the field is too wide, he can't do it. The single-task, that is, the artificial intelligence software for chess is to play chess and can't do anything else.
Economic Observer: In other words, on the premise that these five conditions are met, is the current artificial intelligence capable of part of the job?
Zhang Bo: If your job meets these five conditions, it will definitely be replaced by a computer. The characteristics of work that meets these five conditions are obvious, that is, the four words "do things according to the rules" and do not require flexibility, such as cashiers and cashiers . If your work is flexible and creative, it is absolutely impossible for a computer to completely replace it. Of course, partial replacement is possible, because there must be some simple and repetitive content. If you recognize this article, you will realize that artificial intelligence is still in the early stages of development. It is not as some people estimate that "artificial intelligence technology has fully matured and entered the stage of development and application."
2. "Deep learning technology is close to the ceiling from the perspective of application"
Economic Observer: How should we define the current technical route of deep learning? Is it based on probability?
Zhang Bo: The essence of deep learning today is based on probability statistics. What is called probability statistics? It's not so mysterious. Deep learning is to find those recurring patterns, so if there are too many repetitions, it is considered the law (truth), so a lie is considered truth if it is repeated a thousand times, so why big data sometimes produces very absurd results , Because it doesn’t matter if it’s right or not, as long as it repeats too much, it will follow this rule.
I often say that we have not yet entered the core issues of artificial intelligence. In fact, the core of artificial intelligence is knowledge representation and uncertainty reasoning, because where is the source of human intelligence? In knowledge, experience, and reasoning ability, this is the foundation of human rationality. The artificial intelligence systems formed today are very fragile and vulnerable to attack or deception, require a lot of data, and are unexplainable, and have very serious defects. This defect is essential and caused by the method itself.
Economic Observer: In other words, it cannot be completely solved through improved methods? For example, if we increase the number and complexity of the neural network or increase the magnitude of the data, will it solve its shortcomings?
Zhang Bo: Improvement is impossible. The essence of deep learning is to use the unprocessed data to use the "black box" processing method of probabilistic learning to find its laws. This method itself usually cannot find "meaningful" laws. It only You can find recurring patterns, that is to say, you cannot achieve true intelligence by relying on data alone.
In addition, deep learning is only part of the current artificial intelligence technology. Artificial intelligence has larger and broader areas to study, such as knowledge representation, uncertainty processing, human-computer interaction, and so on. It cannot be said that deep learning is artificial. Intelligence, deep learning is only part of artificial intelligence. Until last year, one-third of the papers exchanged at the Artificial Intelligence Conference were machine learning, and two-thirds were other aspects.
