Deep learning is a research direction in the field of artificial intelligence. It is mainly to learn the internal laws and representation levels of sample data. The information obtained in the learning process is of great help to the interpretation of data such as text, images and sounds. Its ultimate goal is to allow machines to have the ability to analyze and learn like humans, and to recognize data such as text, images, and sounds.
However, there has always been a black box problem in the field of deep learning, that is, the designer of the algorithm does not know how the model draws conclusions and predictions, and the designer cannot explain and be responsible for this result.
Another problem with deep learning is that it is easily disturbed by data "noise". As long as some "noise" is added to the confrontation network, it may make the calculation model wrong. A typical example is that when the training data of an autonomous driving algorithm is maliciously added with "noise" (Editor's note: noise refers to data that cannot be detected by humans but can interfere with the algorithm), the STOP sign may be identified incorrectly. This leads to a car accident.
Algorithm black boxes and unsafe models are easily interfered. These current problems in the field of artificial intelligence raise a question to researchers: in the future, how should we develop artificial intelligence technology so that AI can usher in a quadratic curve growth ?
The answer given by Zhang Bo, the dean of the Tsinghua Institute of Artificial Intelligence and academician of the Chinese Academy of Sciences, and his students is: third-generation artificial intelligence technology.
Recently, at the “2020 Third Generation Artificial Intelligence Industry Forum and RealAI Strategy Conference of Ruilai Wisdom” held in Beijing, Zhang Bo was interviewed by The Paper (www.thepaper.cn). He believed that it is necessary to solve the problem of artificial intelligence. The problems encountered in landing applications need to break existing theories and move towards the third generation of artificial intelligence.
During the World Artificial Intelligence Conference held in Shanghai this year, Zhang Bo explained the third generation of artificial intelligence: The first generation of knowledge-driven AI uses knowledge, algorithms, and computing power to construct AI; the second generation of data Driven AI uses three elements of data, algorithms and computing power to construct AI. Since the first and second generations of AI only simulate human intelligent behavior from one side, they have their own limitations and cannot touch the real human intelligence.
Based on this, Zhang Bo proposed that the development path of the third-generation artificial intelligence is to integrate the first-generation knowledge-driven and the second-generation data-driven artificial intelligence, using knowledge, data, algorithms, and computing power to establish a new Interpretable and robust (transliterated from "robust", academically that means a system or organization has the ability to resist or overcome disadvantages) AI theory and methods, and develop safe, credible, reliable and scalable AI technology.
The development of third-generation AI technology, China and the United States are at the same starting line
In Zhang Bo's view, AI technology based on the third generation of artificial intelligence theory will put China and the United States on the same starting line, and will even be in a leading position as they develop.
As early as 2015, Academician Zhang Bo once proposed the concept of developing "third-generation artificial intelligence". In 2018, Tsinghua University established the Artificial Intelligence Research Institute, with Zhang Bo as the dean of the new research institute, and hired Turing Award winner Academician Yao Qizhi as the chairman of the academic committee. In 2019, Tsinghua University announced the establishment of the Basic Theory Research Center of the Institute of Artificial Intelligence, and appointed Zhu Jun, the Dean and Professor of the Department of Computer Science of Tsinghua University, as the director of the research center.
The task of the newly established basic theory research center is to focus on the shortcomings of current artificial intelligence theories and methods, to overcome the technical limitations of current artificial intelligence, and to establish robustness (robust, refers to the robustness and robustness of the system; robustness is the system The key to survival under abnormal conditions), understandable, data efficient and safe third-generation artificial intelligence basic theory and method system.
In addition to promoting the development of third-generation artificial intelligence in basic theoretical research, in 2018, Zhang Bo and Zhu Jun (RealAI) as the chief scientist were incubated from the Institute of Artificial Intelligence of Tsinghua University as an industry-university-research technology company According to the official introduction of Ruilai Wisdom, the company relies on the third-generation artificial intelligence technology to overcome many shortcomings of general deep learning and fundamentally enhance the reliability, credibility and security of artificial intelligence.
In contrast, the development of third-generation artificial intelligence technology in the United States mainly depends on the DARPA department (US Defense Advanced Research Projects Agency). In 2018, DARPA announced the launch of the AI Next project.
When the project was released, DARPA introduced on its website that starting in the 1990s, DARPA helped lead the second wave of AI machine learning technology, which created a statistical pattern recognizer from a large amount of data. DARPA's funding for natural language understanding, problem solving, navigation, and perception technology has invented autonomous vehicles, personal assistants, and near-natural prostheses, as well as a large number of important and valuable military and commercial applications. However, the second-generation artificial intelligence technology relies on a large amount of high-quality training data, cannot adapt to changing conditions, provides limited performance guarantees, and cannot explain its results to users. The future third-generation AI technology may greatly change military mission planning. Scientists can use machines to conduct research, and humans can interact with machines.
DARPA also gave the focus of the AI Next project, namely the development of new capabilities, high-reliability AI, adversarial AI, high-performance AI, next-generation AI, and artificial intelligence exploration plans.
These points coincide with the development of safe, credible, reliable and scalable AI technology proposed by Zhang Bo.
"Therefore, a very important factor in why we emphasize the third generation of artificial intelligence is that this is the first time in the history of our country that we have been on the same starting line with others to develop the third generation of artificial intelligence. Foreign countries are like us." Zhang Bo told The Paper.
Homogenization of domestic AI companies
In the 2019 ranking of unicorn companies in the global artificial intelligence industry released by Hurun, among the top 40 companies, China accounted for 15 companies and the United States accounted for 20 companies.
In the interview, Zhang Bo believes that this list also illustrates the problems encountered in the implementation of artificial intelligence technology. In terms of quantity, the gap between the AI unicorn companies between China and the United States is not large. But if you carefully study the companies on the list, you will find that the homogeneity of Chinese AI companies is obvious.
"This is not good. If you use the same technology to do the same application scenario, it means that everyone will have this technology, and everyone can use the technology that everyone can think of. The competition must be very fierce." Zhang Bo Say.
In addition, due to the limitations of second-generation artificial intelligence technology, the current domestic AI technology application scenarios are limited, and there may also be security issues.
Insecurity is reflected in two aspects, one is data, and the other is algorithms. "A privacy protection, an intellectual property rights, why the user's face data is used by a company, how to use it, the data is taken away by the company, whether it will cause unfair competition, these all need to be considered and governed." Zhang Bo Say.
Furthermore, due to the fragility of the test environment, many algorithms are easily attacked in the real environment. In the field of autonomous driving, systems are easily attacked and deceived.
This status quo also provides development prospects for the third generation of artificial intelligence technology. Zhang Bo believes that in the future, an algorithm can be explained, people can understand the reasons why machines make predictions, and designers have the means to prevent attacks by others. Only then will artificial intelligence technology usher in better industrial development.
Currently, Ruilai Wisdom, incubated by Tsinghua University, is exploring in this direction. After two years of development, Ruilai Wisdom has recently issued a report card for foreign affairs and released two latest AI products: the privacy protection machine learning platform RealSecure and the artificial intelligence security platform RealSafe2.0 version. The former is based on secure multi-party computing and federated learning. The data security sharing infrastructure created by technologies such as, hidden query, etc., through the opening of data islands, solves the data security risks and privacy leakage issues in the process of corporate data cooperation. The latter is a set of firewalls and antivirus software for AI systems. Detect new security vulnerabilities such as adversarial sample attacks and model backdoor attacks, provide end-to-end solutions for model security evaluation and defense, and greatly improve the security of AI systems.
Finally, Zhang Bo also mentioned the cultivation of talents in the interview. He believes that for the further development of the artificial intelligence industry, it is necessary to develop AI talents with innovative capabilities. "The average level of domestic talents, such as Tsinghua University, is very high in both undergraduates and postgraduates, but the highest level is too far behind others. In our talent training, Einstein has never been shown, nor has it been shown. Spirit. This is what we need to think about in the future." Zhang Bo said.