The historical change of AI research direction from machine learning human behavior to human research machine behavior 【Full Text】

A few days ago, a security research and development network led by the MIT Media Lab, a research group with multiple researchers from companies such as Harvard, Yale, Mapu, and Microsoft, Google, and Facebook published an article in Nature with " A review article titled "Machine behaviour", proclaiming the emergence of the new discipline "Machine Behavior" that spans multiple research fields.

"Uncontrollability" of machine algorithms

At the beginning of the rise of AI, robots, etc., due to the asymmetry of information and people's cognition, there is a fear that "robots will generate self-awareness" in society. People worry that when the robot has self-awareness, it will in turn control humans and endanger human safety.

With the maturity of machine algorithms, the popularization of products and the improvement of people's awareness in recent years, people's fears have decreased. However, more and more scientists have begun to discover that there are some "uncontrollable" factors in machine algorithms, and are actively looking for solutions. For example, Li Feifei announced the establishment of Stanford "Human-oriented AI Research Institute", the establishment of robot behavior standards in the UK, and the leadership of MIT. The new discipline of "machine behavior" and so on.

Among them, changes in the environment and human decision-making are constantly affecting the behavior of the machine. It is understood that machines have inherent mechanisms for generating behaviors. These behaviors acquire information, develop in the interaction with the environment, and produce corresponding functions, resulting in specific machines undergoing different changes in their corresponding environments. For example, smart products for the elderly, the intelligent system detects and analyzes the elderly's actions in real time to determine whether the elderly are eating normally, taking medicine, keeping a low amount of exercise, and whether there are abnormal movements (such as falling), so as to promptly remind and ensure The quality of life of the elderly will not be reduced by living alone; the hospital's intelligent rehabilitation training, its intelligent robot recognizes the degree of normative movement behavior, and evaluates the degree of recovery to provide better rehabilitation guidance.

Although machines and animals have essential differences in attributes, experts believe that intelligent machine behavior research can be helped by the study of animal behavior, through the four dimensions of the function, principle, development and evolution history of a behavior. The basis for the establishment of machine behavior.

The "Machine behaviour" article mentions that according to the integration and influence of artificial intelligence systems, the research scope of machine behavior includes: single machine behavior, emphasizing the study of the algorithm itself; collective machine behavior, emphasizing the study of the interaction between machines; Mixed human-machine behavior emphasizes research on the interaction between machines and people. Machine behaviors and human behaviors can both shape each other, or they can interact collaboratively.

From machine learning human behavior to human research machine behavior

Since the rise of artificial intelligence research, the industry's research is mainly to allow machines to understand human social life through big data computing, deep learning and other means to assist human work or life.

With the development of artificial intelligence, AI, big data, robots, etc. formed by relying on algorithms are gradually integrated into our lives and become more and more important roles in human society. However, due to the complexity and extensiveness of intelligent computing, there are more and more "output difference" problems in intelligent algorithms. The potential safety hazards such as intelligent driving judgment errors cause accidents, and service robots say inappropriate words. It is difficult to understand their behavior by relying on data analysis. This makes people have to explore the reasons.

Machine behaviour is mainly to study the behavior exhibited by intelligent machines. The emergence of the concept of "machine behavior" also symbolizes a major turning point in human research-from letting machines study human behavior to humans studying machine behavior.

It is understood that the complexity of individual artificial intelligence bodies is already very high and continues to grow. Although the code for structuring them and the training used to train the model can be concise, the trained model will not be so simple at all, which is often This resulted in the generation of "black boxes". The artificial intelligence body accepts the input and outputs it, but even under the current situation that some application scenarios of "interpretability" have progressed, the artificial intelligence body actually generates these output processes, which is difficult for the scientist who constructed them to explain.

In addition, the source codes, models, and data sets that are frequently used in society are actually copyrighted. This leads to that in many scenarios, people can only observe the input and output of the system, and cannot observe the machine during the input process. What kind of changes have taken place makes it more difficult for people to understand "machine behavior".

Humans can create intelligent machines, and intelligent machines can also change human behavior, but from the perspective of the difference between input and output and the opacity of machine learning, it is currently unknown whether the impact is positive or negative. Only through systematic research and understanding of “machine behavior” can we help people understand how ubiquitous algorithmic systems work in society, supervise their possible subsequent consequences, and then carry out trade-off technological innovations to make our Social life is safer and safer.