Artificial intelligence can be divided into special artificial intelligence and general artificial intelligence. The current progress is mainly achieved by special artificial intelligence. The research and application of general intelligent systems still have a long way to go.
The Alpha dog defeats the human Go champion; artificial intelligence surpasses the human level in large-scale face recognition; the accuracy of the speech recognition system is comparable to that of professional stenographers... In recent years, the development level of artificial intelligence has been remarkable.
Why should we strengthen the research on the basic theory of artificial intelligence?
Academician Tan Tieniu, a researcher at the Institute of Automation of the Chinese Academy of Sciences, believes that artificial intelligence can be divided into specialized artificial intelligence and general artificial intelligence. The current progress is mainly achieved by special artificial intelligence. The truly complete artificial intelligence system is a general-purpose intelligent system that can be used in the same way as the human brain. However, the current artificial intelligence system has IQ, no emotional intelligence, will not calculate "calculation", there is no general talent. The research and application of general intelligent systems still has a long way to go, and the overall development level of artificial intelligence is still in its infancy. "The basic theory of artificial intelligence frontier is the cornerstone of artificial intelligence technology breakthrough, industry innovation, and industrialization. In order to achieve the final right to speak, China must make a major breakthrough in the basic theory of artificial intelligence and cutting-edge technology."
The experts interviewed pointed out that the space for exploring the basic theories of artificial intelligence is very huge. Take the current basic theory of artificial intelligence, the deep learning method, as an example, it is neither perfect nor more than the basic theory of artificial intelligence.
“Deep learning methods have limitations.” Song Jiqiang, dean of Intel China Research Institute, said that deep learning can recognize faces, but can’t predict the emotional relationship with another person through one’s speech because it lacks this. Aspects of knowledge input.
Wu Gansha, the person in charge of the technology, said that if the data is not accurate, the data set is biased or even “fighting” the input of false data, deep learning may go wrong. It is obviously a panda. As long as it changes hundreds of pixels, deep learning may identify it as a seal.
Wu Fei, deputy dean of the School of Computer Science at Zhejiang University and director of the Institute of Artificial Intelligence, believes that from the data-driven field of artificial intelligence to human universal intelligence, there are still unknown areas of neuroscience, cognitive science and even new mathematical models that need to cross. The road is still far away.
"Scientists should be encouraged to target the frontiers of artificial intelligence disciplines, carry out leading original scientific research, and focus on the major basic scientific issues in the field of artificial intelligence through the intersection of artificial intelligence and brain cognition, neuroscience, psychology and other disciplines. The formation of the original theoretical system of artificial intelligence with international influence will provide theoretical support for the construction of China‘s independent and controllable artificial intelligence technology innovation ecology." Tan Tieniu said.
Look at the key core technologies -
Artificial intelligence can be divided into infrastructure layer, enabling layer and application layer. There are many core technologies on all three levels, and there are many unresolved technical bottlenecks.
In recent years, China‘s artificial intelligence technology has made great progress, but some key core technologies still need to be overcome.
In Wu Gansha‘s view, artificial intelligence can be divided into three layers: infrastructure layer, enabling layer and application layer. Among them, the infrastructure layer refers to the basic theoretical algorithm and computing power; the enabling layer involves specific application scenarios, such as speech recognition, computer vision, expert systems, gaming systems and robots; the application layer refers to unmanned driving, intelligent medical, etc. Integrated scene.
"There are many core technologies at all levels, and there are many unresolved technical bottlenecks." He said that from the perspective of infrastructure, there are still a lot of breakthroughs in cutting-edge basic theories and algorithms, and hardware devices such as chips need further Localization. From the perspective of the empowerment layer, the artificial intelligence related to speech recognition has made great progress overall, but some intelligent machines sometimes translate language that is not like “human language” and lacks a deep understanding of semantics; industrial robots can handle several The problem of the link, but can not achieve precise control through muscle and nerve sensitive manipulation in the final assembly. From the perspective of the application layer, unmanned driving is an open, dynamic and uncertain environment. Humans cannot “feed” all the traffic scenes in the machine world. Unmanned vehicles may not be able to handle certain situations. If the parking lot does not have a map, the unmanned vehicle may not be able to read the marking line, so that it is impossible to find a parking space or a road leading to the exit.
Experts pointed out that in order to overcome the breakthrough of the frontier basic theory, we should pay attention to the basic hardware research and development, support system research and development, ecological construction and research and development mentality adjustment.
Zhou Zhihua, director of the Department of Computer Science at Nanjing University and dean of the Institute of Artificial Intelligence, said that almost all smart applications are now inseparable from GPUs. Many companies directly deploy smart applications in TensorFlow (symbol mathematics based on data stream programming). On systems such as systems, this is at risk of being "card necked" in the future. Domestic efforts should be made to develop the basic hardware and support systems for machine learning, and to produce replacements for GPUs and TensorFlow. On the other hand, both GPU and TensorFlow are based on deep neural network models. If breakthroughs are made in the deep learning of non-neural network models, the “monopoly” of these basic hardware and systems based on deep neural network models will naturally disappear.