Guo Yike, vice president of Hong Kong Baptist University: when we turn on the mobile phone to read the news we want to know, is the push that slips past our eyes is the news that we are really interested in? When we click on the webpage to look for the goods we want to buy, are the treasures that pop into our eyes are the products we really like? When we log in to the program and want to make an appointment with a Uber, the system will display the price Is Ge really what we need to pay? For these common scenes in life, most of us are used to it. Unconsciously, artificial intelligence has invaded our lives. Whenever we start a long distance conversation with the Internet, we cant tell whether the results we want to get is whether they are given by humans or by machines. We are immersed in the bubble created by artificial intelligence for us in the life of artificial intelligence. We hope to return to the original nature, pay attention to the ethical issues of AI under the philosophical proposition, and bring cool thinking for the future development of AI.
Xi Qing: with the wider application of artificial intelligence, there are more and more controversial issues. Is the principle behind intelligent recommendation reasonable? Does algorithmic pricing in financial markets hide malicious competition and price manipulation? Is there a legal basis for the deprivation of rights in the social credit system? Does the emergence of sex and chat robots change human natural emotions? These controversial issues bring a new perspective to the research of artificial intelligence. Are these ethical issues?
Guo Yike: in fact, the so-called intelligent ethics is the form and change of jurisprudence and organization in the society of man-machine coexistence. As we all know, the development of artificial intelligence is inseparable from big data, model algorithm and computing power. High quality data resources, advanced learning algorithms and the computing power to support the implementation of the algorithm have helped the development of artificial intelligence. Good data ecological environment and social ethics are the two important factors for the smooth development of artificial intelligence. People constantly ask machines to do more, but often ignore whether machines do right or not. This is why the ethics of artificial intelligence has not been paid more attention. When people endow machine intelligence, they should also understand and examine the behavior of machines. This is the research category of intelligent ethics. The ethics of artificial intelligence is not only a problem of technology and science, but also a problem of philosophy. Different understanding and values may not get the same answer, so the research of AI ethics will be an open topic with the development of AI in the future.
To explore the origin of the ethical problems of artificial intelligence, we start with machine learning. If we understand machine learning from the perspective of cognitive science, it is very similar to the logic and process of human brain learning. We regard the knowledge in the human brain as a model, and the cognitive model of the human brain is the abstraction of the view of the world and the observation results. The human brain obtains information and results through observation. When the observation results are consistent with the cognitive model of the human brain, the human brain will not do any behavior except to produce some pleasant dopamine; when the observation result is inconsistent with the human brain model, the human brain will react. It can be very rational and think that there is something wrong with its cognition, so it will make a response to the cognitive model in the brain Change, or it is very confident, fully believe in its own model, so, put into action, to change the world. The human brain changes cognition is learning, changing the world is the action under the guidance of decision-making. This is the cognitive understanding of learning.
The same is true of machine learning. In machine learning, models are expressed in the form of functions. The purpose of machine learning is to find the form of a function and the parameter of the function. The machine uses this parameter to compare the observed results, so as to minimize the error. In machine learning, whether it is supervised learning, unsupervised learning, reinforcement learning, there is only one goal, to find a model consistent with observation. Looking at the process of machine learning, we will find a series of fundamental problems. First of all, how to express peoples cognitive purpose to machines. What people tell machines to do is expressed by utility function. It is very difficult to define peoples cognitive purpose with utility function. We should not only achieve the best results, but also meet the requirements of social ethics. Secondly, machine learning is also a process of utility function optimization, which is very difficult for us to understand. Optimization must adjust many parameters, change some controllable variables, just like alchemy, the process of parameter adjustment will lead to many problems.
Xi Qing: while giving machines super capabilities, people enjoy the benefits brought by machine behavior, and at the same time have many doubts about machine behavior, including whether the generation algorithm in news writing is reliable? Will the trial machine help the trial bias? Is the machine correct in medical diagnosis?
Understanding machine behavior is an important issue in artificial intelligence research, and also the main direction of artificial intelligence research in the future. Artificial intelligence in the future will not require machines to do more, but to do them right. What is right is not only a question of philosophy, but also a question of values. How to express the purpose of the right behavior to the machine, how to judge whether the machine is doing right or not, how to understand the reason why the machine does the right thing, and how to explain and verify the machine behavior are all important issues to understand the machine behavior. Expressing the right purpose to the machine is the main research direction of the ethics of artificial intelligence. When we use utility function to express learning purpose to machine, the correctness of function expression and the correctness of verification function need to be considered; whose benefit maximizes? Here is the so-called agent principle: if you find a lawyer to work for you, you need to make sure whether the lawyer works for you or makes your business more complicated, so that he can make money There is a principle of unpredictability - that is, the problem of good intentions doing bad things needs serious consideration. In the final analysis, the ethics, rationality and fairness of expressing objective functions to machines are all problems of expression. To judge whether the machine is right or wrong, how to verify it? In the process of verification, what is the best? Whether the optimal is reasonable? What kind of nature should artificial intelligence system have to win trust? These are also ethical issues of artificial intelligence that need to be further studied. Future learning should increase the understanding of the causality of machine behavior. Our ongoing research on the semantic interpretation of hidden space is related to this. When the machine is learning from zero samples, we interpret the semantics of the hidden space and edit some features into vectors. When an unknown object appears, the machine can get the vector by coding, and then deduce the features from the particle. We plan to enlarge the concept of hidden space and map a probability model into the concept space to explain many deep-seated methods.
Xi Qing: you introduced the ethics of artificial intelligence. We found that it is far from the true Turing era. Because when we ask questions in the dark room where people and machines coexist, we cant tell whether the answers come from people or machines. We can ask, please tell me, how did you get the answer to this question? people can answer this question, but machines must be confused. In the future, how do you understand the world of man-machine coexistence?
Guo Yike: people can answer this question, but machines must be at a loss. However, for the world of man-machine coexistence, we have no doubt. In fact, we are more and more used to symbiosis with machines in this new dual world. As we interact with machines, we also try to understand the behavior of machines. Only when we understand the behavior of machines, can we believe in the behaviors of machines and the machines that we get along with day and night. The machine must understand the human behavior. The interpretation, understanding and verification of machine behavior, as well as the ethics of machine behavior, will become the core focus of artificial intelligence research, and also the difficulties and pain points that scientists strive to overcome. It is too early for us to talk about the rule of artificial intelligence in the future before we solve and deeply understand these problems.
Artificial intelligence is not a myth, but a real algorithm. The research of artificial intelligence reflects the continuous pursuit of the future and the continuous exploration of the unknown. Artificial intelligence, we can neither despise, nor fear, nor avoid. We should develop solid technology with rich ideas through down-to-earth efforts and hard work, and build a New Dualistic world with the unique tolerance and understanding of Eastern civilization. The coexistence of intelligent machines and human beings is the future development direction of artificial intelligence and the blueprint of our future society.
This article source: Xinhuanet editor in charge: Wang Fengzhi_ NT2541