The hottest AI cancer diagnosis expert says it's u

  • Detail

Artificial intelligence to diagnose cancer? Experts say it's unreliable


Wang Xiaotao source: China Economic Herald word count: 2206

China Economic Herald | Wang Xiaotao

last October, when "Sofia" won the citizenship of Saudi Arabia, this beautiful robot was a hit. However, Professor Zhang cymbal, academician of the Chinese Academy of Sciences and Dean of the Institute of artificial intelligence of Tsinghua University, disagreed. At the 2018 World robotics conference · Youth Innovation and Entrepreneurship Forum held a few days ago, he revealed that: "Sophia talks freely on TV and performs very well. The main reason is that the questions it answers are all raised in advance, so it's a little fishy here."

Zhang Dan, an academician of the Canadian Academy of engineering and professor of York University, Canada, also believes that artificial intelligence is developing very fast, but people should think calmly. In fact, there are many things exaggerated. "The measurement principle of shore hardness tester is that robots such as Sofia are fake. In fact, the dialogue between humans and robots is impossible." He said

"Americans accuse Hansen, who studies Sophia, of being a liar, but it is essentially an entertainment product", Zhang cymbal said with this as an example, "I want to tell you that we should not set too high an index when we are doing industry. You can't actually do too high an index. Like Sophia, it can only be a lie."

the development of robot is inseparable from artificial intelligence. It is often said that the development of artificial intelligence depends on three factors: algorithm, data and computing power. However, academician Zhang cymbal believes that the most critical thing is not these three, but "application scenarios". He said: "in fact, all enterprises have more or less algorithms, data and computing power, but the most important reason why some enterprises succeed and others fail is that the application scenarios are not selected properly, which is the biggest problem."

as we all know, in the field of artificial intelligence health care, many companies are conducting research on cancer diagnosis, and some research results claim that the recognition rate of artificial intelligence is close to or even higher than that of human beings. With such good technology, will it be successful to set up an enterprise or apply it to medical treatment? Can it be used to play the role of a public service platform for military civilian integration? The results were disappointing. Zhang cymbal said that even Watson, which IBM is proud of, gave "many unsafe and incorrect treatment opinions" in practice, and even "prescribed drugs that are easy to cause bleeding for cancer patients with hemorrhagic disease, which can cause death in severe cases."

professor wangtianmiao, honorary director of Robotics Research Institute of Beijing University of Aeronautics and Astronautics, said on another occasion: "Watson has spent tens of billions and is currently at a low ebb in medical treatment. When judging cancer and surgery plans, doctors are difficult to understand its judgments and plans. When asked, they can't answer, so many cooperative hospitals quit."

in fact, Watson is undoubtedly the best example of the unexplainability of artificial intelligence systems. Zhang cymbal said that in addition to the privacy and protection of personal data and the reform of rules and regulations, the first problem to be solved in the application of artificial intelligence in health care is the interpretability and robustness of artificial intelligence systems. "In fact, there is such a problem. If the intelligent image recognition tells the doctor that the patient has cancer, does the doctor believe it? If it is diagnosed incorrectly, whose? So the enterprise claims that its recognition rate exceeds that of people, but when it is used by the doctor, the doctor is not at ease." Academician Zhang cymbal believes that this is actually one of the most important problems in deep learning, that is, unexplainability. "Moreover, its robustness is very poor, and it is easy to be cheated. You can design a noise for it, and it can be recognized as anything. In other words, it basically does not know what cancer is. It does not recognize cancer according to the characteristics extracted by doctors, but finds the differences from a large amount of data," Zhang cymbal emphasized, "The diagnosis made by the machine is fundamentally different from the diagnosis made by the doctor. Because of this difference, doctors dare not use it, although your recognition rate may be higher than that of people."

the temperature is about 23 degrees Celsius. Now people hold great esteem for deep learning, but academician Zhang cymbal said: "the results of deep learning seem to have a higher recognition rate than people, but the principle it uses is different from people. Therefore, people are not at ease when using such a machine to help people make decisions, see a doctor and make other decisions." Therefore, he believes that in addition to the three elements of data, algorithm and computing power, knowledge is a more important factor for the success of artificial intelligence. "At a knowledge map conference just concluded in Tianjin, 1/3 of the more than 800 participants came from enterprises. Those enterprises were very successful. They relied not on data, but on knowledge. Now many enterprises have realized the importance of knowledge earlier than schools and research units." Academician Zhang cymbal believes that the knowledge and experience of doctors looking at pictures must be added to the in-depth learning method for medical image recognition. Otherwise, the results obtained by relying on data alone cannot interact with doctors, and doctors cannot believe it, and naturally they will not use it

Zhang Dan said that it is impossible for deep learning to replace all technologies. "Deep learning has indeed made great contributions to face recognition and language synthesis, but its deep learning is actually summarized based on a large number of data laws, which is completely different from the process of human learning. Therefore, it is impossible for deep learning to completely replace people"

in what application scenarios is AI more likely to succeed? Academician Zhang cymbal can be summarized into five points: first, to master rich data or knowledge; Second, complete information; Third, certainty information; Fourth, static and structured environment; Fifth, limited fields or single tasks

based on this, Zhang cymbal believes that those jobs that are easily replaced by machines mainly do normative daily affairs, such as cashier, retail waiter, office clerk, hotel waiter, bookkeeper, accountant, audit, etc., which is to act according to the rules. Jobs that are not easily replaced by machines do not follow the rules, such as dynamic changing environment, incomplete information, uncertainty, multi domain and multi task, etc

similarly, considering the dynamic and uncertain environment, academician Zhang cymbal believes that in the streets like Beijing, driverless driving is simply impossible to achieve in the short term. "The method of in-depth learning is to learn from the sample. It is impossible to learn all the situations, because you can't estimate what a pedestrian will do and how he will cross the road. Therefore, it is impossible to rely on it to deal with emergencies. The industry must consider the difficulty of this problem." He said

for the industrialization of artificial intelligence, academician Zhang cymbal said that artificial intelligence has just started, and a large number of research tasks need to be done. It is necessary to establish a good "government, industry, University and research" cooperation mechanism. 2. Protection and maintenance of bellows ring stiffness testing machine. At the same time, it must be combined with the real economy to create value, because artificial intelligence is an application-oriented discipline. It is necessary to solve practical problems, especially combine with local reality, and select appropriate application scenarios for development

Copyright © 2011 JIN SHI