"Please give AI some tolerance"

Behind the misunderstanding of AI, the talent gap has reached 300,000.

”

Author | Gao Xiusong

Edit | Yu fast

"To be honest, I don't like the term "artificial mental retardation."

In a chat with Leifeng.com, an algorithm engineer engaged in computer vision said many times that he had hated the word for a long time, almost instinctively disgusted, even if it was just a joke, in his opinion it was a kind of ridicule.

This kind of ridicule is like a passerby sneering at a child who has just learned to climb: this child is so stupid that he can't even walk.

He even confessed that if a colleague around him used the word self-deprecating, he would deliberately keep a distance from him, because this self-deprecation was really "disrespectful" to his work and professional knowledge.

There are not a few engineers with his technical habits, and among the many practitioners asked by Leifeng Network, they have expressed similar views: usually when asked about the level of artificial intelligence, similar expressions are based on "weak artificial intelligence".

A business commissioner in charge of brand communication and public relations revealed that if words such as "artificial mental retardation" are used in external communications, they are reported or discovered by the company, "directly affecting performance appraisal", because such unprofessional expressions are likely to lead to negative communication effects.

In the conversation with these people, Lei Feng network found that in the AI circle, practitioners have a clear understanding of AI, and when publicizing, the negative expression of AI is more rigorous.

However, outside the circle, a series of AI accidents have caused the public to have many doubts about the true ability of AI, and the rhetoric about artificial intelligence becoming artificially retarded is rampant, and the voice of artificial intelligence is often reported in the news.

On the surface, this is just a public opinion controversy about AI. However, its essence is the competition between enterprises and the public for the right to speak about AI, and will directly affect the promotion, landing and application of AI.

"If the public cannot form an effective understanding of new technologies, then the promotion of new technologies is very slow." A graduate student at a media university said that the public's ability to accept new technologies is gradually progressing, this process is easily affected by public opinion, and negative public opinion has a "blasting effect", which may directly destroy the "trust foundation" previously established.

For example, autonomous driving, the public's trust in it is very weak, and after many accidents, this trust has actually been exhausted.

Relevant research reports show that the activation rate of Tesla FSD, the brother of automatic driving, is less than 10% in China, and even a considerable number of people have not opened AP services, even in the open population, few people will use AP functions.

Although this phenomenon has its objective reasons (such as insufficient roadside data and limited algorithm capabilities), from the perspective of public opinion dissemination, a mistake in automatic driving is more serious than the ten car accidents of traditional cars, which also hinders the further landing of automatic driving.

So, how to establish an effective understanding of AI for the public and promote the faster and more extensive landing of AI?

After the interview, Leifeng Network believes that media reporting, enterprise external publicity, and public knowledge popularization education are the three most important ways. The various "cognitive education" around the public is also destined to be a protracted "battle".

1

"I have never been to the sea, I don't know the depth of the water"

There is an interesting paradox in the application of artificial intelligence: when an AI technology is already very popular, people generally do not think that it is AI.

For example, in the 1980s and 1990s, a black-and-white TELEVISION may have been an epoch-making symbol that required manual FM, but now that remote-controlled color televisions have become standard, people don't think it's smart. For example, the community parking lot through the license plate recognition in and out, brush face into the community, etc., in recent years began to popularize, but people rarely associate it with AI, even if it actually uses a variety of recognition algorithms, chips and so on.

In the public perception, artificial intelligence should naturally reach the level of robots in movies, or think and act like humans.

"The public is sometimes overly optimistic about artificial intelligence, even overestimating." Yang Li, associate professor of the School of Information of China Jiliang University and head of the artificial intelligence major, believes that as a new technology moving towards society, people's understanding of AI is not comprehensive, and they believe that AI should be omnipotent, which is not consistent with reality.

In the view of Leifeng Network, the public's cognition of artificial intelligence is relatively shallow, which is mainly manifested in two aspects:

Do not understand, do not know, for what is artificial intelligence, there is no direct impression;

There is a certain understanding, but can not understand the deep logic (principle structure) of AI.

This shallow cognition is easily induced, and under some marginal publicity, the ability of AI itself is exaggerated, and the public has blind "confidence" or overestimation of AI.

"The layman looks at the bustle, and the insider looks at the doorway."

Yang Li said that taking face recognition as an example, 5 years ago, people may feel very mysterious and advanced, but after the popularization of consumer electronics, many people feel that face recognition is no longer difficult. When he lectured the students on face recognition, the students felt that this was already a very mature technology, "not new, not difficult." ”

But in fact, face recognition is still a long way from being highly intelligent, and in many complex scenarios, it is difficult to capture effective face information. Moreover, face recognition works well in small-scale (small database) scenarios, but when the database is very large, the accuracy of recognition is not so high.

"Due to the lack of professional knowledge, it is easy for the public to simplify complex problems, but people engaged in AI research are very cautious about this, ordinary people feel that simple technology, practitioners may feel that 'this can't be done, that can't be done', in short, it is the feeling of looking at the mountain and running a dead horse."

Leifeng network found that due to the lack of professional general education, the public's understanding channels for artificial intelligence are relatively single, most of them are through media reports, corporate publicity and other two ways to touch AI, only a small number of people will spontaneously study relevant books, learn courses to enhance understanding.

From the perspective of communication, if the audience's access to information is limited, then the controller of the information channel will have the "control" of information transmission, forming a "public opinion monopoly" situation, and the information is easily "distorted" after multiple dissemination.

In fact, this kind of "distortion" is inevitable. In the process of AI dissemination, two major groups have been formed, inside and outside the circle, because artificial intelligence itself belongs to a higher threshold of professionalism, the connection between the circle (enterprise) and the outside circle (general audience) is mainly achieved through the media.

However, the problem with media publicity is that many practitioners are either born in the class or cross-border transformation, and only a few media people really understand AI. And with the changes of big data and Internet technology, the media itself has further sunk to various platforms, creating countless self-media, forming a situation in which the media industry is uneven. In a traffic-oriented environment, various news reports emerge in an endless stream, and this kind of information has a "amplification effect" (such as the headline is too shocking), so that there is an "error" between the information received by the public and the actual information.

At the time when artificial intelligence was the hottest, many AI companies in order to take financing and play popularity, have put advertisements, soft texts, and publicize products, resulting in the illusion that artificial intelligence has been able to land on a large scale. Later, AI was cold, and the public's ridicule of AI can be seen as a kind of "anti-phagocytosis" that was too violent in the early stage of propaganda.

Of course, the circle also noted the limitations of mass media, many companies have opened up publicity channels on important social platforms, but due to content differences (such as too vertical, product promotion) or channel differences, and do not meet the C-end attributes, most AI companies can not directly establish an effective connection with the public.

Therefore, under the communication chain of "enterprise-media-public", due to the mechanism defects of mass media itself, it is difficult for the public to establish an effective understanding of AI in uneven information. However, enterprises have to rely on mass media to promote AI, and this internal contradiction is an important reason for the "cognitive difference" between AI inside and outside the circle.

"At the end of the day, there are too few AI talents." In Yang Li's view, talent is the core force to promote the development of the industry, the current AI is in the climbing stage, the problem of technology itself is the fundamental factor that causes the public to question AI, and the dissemination of public opinion has exacerbated this impact to a certain extent.

Whether it is the in-depth development of AI or horizontal dissemination, only AI talents can give AI a "correct name", but the situation at this stage is that domestic AI talents are extremely scarce.

2

"Ride the wind and waves, talent first"

"There's really too little application talent." Yang Li lamented that when AI moves from the castle in the air to the field, there are "really not many people" who understand technology and industry.

In the Ministry of Industry and Information Technology's "Artificial Intelligence Industry Talent Development Report (2019-2020)" (hereinafter referred to as the "Report"), it is estimated that the effective talent gap in the mainland artificial intelligence industry will reach 300,000, which is only two years ago. In fact, in the past two years, according to Leifeng Network observation, the demand for talents in AI companies has continued to be strong, and the application talent gap in the entire AI industry has further widened.

As a technology/knowledge-intensive industry, AI has a high threshold for talent access and attaches great importance to academic qualifications and work experience.

According to the report, only 11.9% of the positions released by AI companies in 2019 accepted a college degree; only 5.4% of the positions accepted job seekers with less than 1 year of work experience; and only 3.3% of the positions that provided fresh graduates were accepted.

This means to engage in the AI industry, basically requires a bachelor's degree, at the same time, because most AI companies lack manpower, funds and motivation to train fresh graduates (at least more than one year), the demand for graduates is not strong, and prefer those with knowledge reserves and practical experience of talents, this "new" nature of the recruitment demand, but also aggravated the shortage of talents.

In addition, AI has strong professional requirements for talents, especially in positions such as algorithm research and application development, and more than 60% of the positions require computer and mathematics-related professional backgrounds.

Under the constraints of various linear conditions, the already shortage of AI talents is more "tight".

An AI startup HR told Leifeng Network that it is a very difficult thing to recruit people, "professional, school, work experience screening down, there are very few qualified people, plus the company wants to come in and produce people immediately, but also consider the salary of these factors, excellent talents are difficult to recruit; and if you go to school recruitment, excellent graduates are signed by the Internet and star AI companies early, and the rest are more favored by large companies." Filter to filter, there are really not many choices. ”

In addition to the lack of application-oriented talents integrated with the industry, in Yang Li's observation, another talent gap in AI is theoretical research talents who can "lay down their hearts to do basic work".

According to the "2022 Artificial Intelligence Report" released by Stanford, although the mainland ranks first in the world in the number of CITations of AI journal papers, the number of conference papers published, and the number of AI patent applications, it lags far behind Europe and the United States in the number of CITations of AI conference papers. Moreover, some innovative basic theories and cutting-edge scientific and technological research are still mainly in Europe and the United States.

"A lot of the basic theories of artificial intelligence are proposed by foreigners/institutions, such as deep learning, which is now hot."

Yang Li said that this has a lot to do with the late start of artificial intelligence in the mainland, to make up for such a gap, in addition to strengthening the funds and talent investment in basic theoretical research, we should also establish a standard AI talent training system to provide a steady stream of talent vitality for AI research.

"Schools are the cradle of cultivating talents, and the ideal situation is that some students engage in theoretical research after graduation, and more graduates enter the industry to promote the landing of AI through the linkage of production, education and research."

Leifeng Network learned that the current mainland artificial intelligence industry has initially formed a "government-industry-university-research integration" talent training ecosystem, but it is still in its infancy. In 2019, the artificial intelligence major was officially approved to be included in the list of undergraduate majors, and many domestic universities began to build artificial intelligence colleges (research institutes) by themselves or with enterprises, and opened AI majors.

However, for how to cultivate professional AI talents, major universities are also groping, and have not yet formed an effective paradigm.

3

"Teaching according to aptitude, stimulating interest"

In 2019, the domestic artificial intelligence major was officially approved and included in the list of undergraduate majors, but the opening of majors needs to go through the process of curriculum construction, experimental conditions, professional declaration, etc., and most schools have only begun to formally enroll students in the past two years.

In other words, it will take about one to two years before the first batch of AI undergraduates graduate.

How to cultivate this batch of new students into talents to fill the current talent gap is not an easy task. In addition, the first batch of future graduates, whether their comprehensive ability to meet the standard is also of great symbolic significance.

"On the one hand, the content of artificial intelligence majors is very difficult, and many courses that were only opened at the graduate level in the past are now put to the undergraduate level to learn, which is a pressure on students, and it also brings challenges to teachers' teaching methods and skills; on the other hand, how to combine talent training with social needs so that students can apply what they have learned is also a difficult point."

As a senior scholar in the field of artificial intelligence, Yang Li has not only had in-depth research and thinking on AI in his coaching career for many years, but also explored some "methodologies" on cultivating AI talents.

"First of all, we must respect the law of learning." Yang Li told LeiFeng Network that AI itself has higher requirements for practical ability, which cannot copy the training mode of traditional disciplines, that is, the first and second years of college focus on theory, and the third and fourth years focus on majors. Instead, theory and practice should be used together, first learn, then practice, learn in practice, and then show a "spiral".

In terms of specific measures, he said that through the establishment of a "science and technology innovation group" model, students can be encouraged to participate in various learning competitions and research topics in a team way.

The advantage of this group model is that the group covers all students, through teamwork, forms an internal learning atmosphere of mutual help and mutual assistance, so that members can participate in the practice and become an "interest group"; and the duration of the group covers the student's entire university career, and all members can share the "benefit results". At the same time, the group members help each other, which can also reduce the pressure on the teacher to some extent.

"Secondly, it is necessary to teach according to the aptitude and stimulate students' desire to learn and explore AI."

Yang Li said that students' interest in learning AI also shows a clear "law of two and eight", that is, 20% of students have a strong desire for knowledge, while 80% of students have a general interest.

"For the 20% of students, you only need to tell him how to do the best, and tell him the things and details that need to be paid attention to in the process, and the rest do not need to pay too much attention; for the 80% of the students, their interest is not so high, they need more detailed guidance, and they need to be accompanied by some "mandatory assignment", such as direct assignment of tasks for them to participate."

"In addition, through the incentive mechanism to stimulate students' creative inspiration."

For example, in the design of the curriculum, innovation is incorporated into the scoring criteria, and the course performance is used to drive students to innovate.

For example, in a case, if a student just follows the steps listed by the teacher, the highest grade may be just right, and the rest of the score is all about personal creativity and play.

"Most students need some push from their teachers, and grades are the best motivator." Yang Li said that in order to get higher grades, students have to "think more" rather than perfunctory, and the final works "often have many unexpected highlights".

"Finally, a positive cycle of positive interaction should be formed between teachers and students."

A common problem with undergraduate teaching is that the interaction between students and teachers is weak, or exists only in the classroom, and there are very few extracurricular connections, and it is not uncommon for "teachers and students in class and passers-by in class".

In Yang Li's view, if the teacher only treats teaching as a work task to complete, then the students will also adopt a coping attitude. On the contrary, if the teacher is responsible, the students will also be affected by their "example" and more enterprising.

Therefore, teachers can communicate with students through projects, online and offline interactions, etc., to understand the needs of students, to give feedback to their own teaching work, and this feedback will eventually reach students through teaching, forming a "win-win situation for teachers and students".

In addition to the methodology of cultivating AI talents, Yang Li also pointed out that the cultivation of artificial intelligence professionals needs to break the "only graduate school theory".

"To study artificial intelligence, you must go to graduate school, and there is no future without going to graduate school."

Many people hold such views, but Yang Li firmly opposes them. He believes that the original courses of many graduate students have been delegated to undergraduate learning, after the undergraduate stage of talent training into a system, students' theoretical and practical ability will be able to meet the basic needs of the AI industry, blindly pursuing graduate education, will only cause the AI circle to become more and more volume, will not help alleviate the shortage of talents in the industry.

"Of course, graduate education is also very important, but the cultivation of graduate talents may be more inclined to the basic theory, and the large-scale landing of AI needs more application-oriented talents to promote."

For example, many traditional manufacturing industries have introduced artificial intelligence, such as robotic arms, automated production equipment, etc., but due to the lack of application-oriented talents, the equipment bought back by enterprises does not know how to use, nor how to maximize benefits, let alone operation and maintenance.

Such a position does not require practitioners to have a very deep theoretical foundation, but talents who have an AI foundation and understand the industry. In the process of intelligent upgrading of traditional industries, the similar talent gap is very large.

"In fact, when AI goes to all walks of life and lands, the demand for talents will also change, and at the undergraduate level, through theoretical learning and social practice related to the profession, excellent talents can also be cultivated."

4

"The Bumpy Road to the Great Era of AI"

At the just-concluded Winter Olympics, Professor Yang Li led his team to make an intelligent assistive technology, which can review and analyze the actions of the players through video, and give the referees a reference.

Although it is only a relatively simple behavior recognition, the model is not exquisite, and there are many AI companies in the market that have the ability to develop the technology. But it is gratifying that as soon as this project was proposed, the students enthusiastically participated, under the guidance of the tutor, step by step mining data, labeling, modeling, training, testing, the whole process lasted for two weeks, most of the work was completed by the students, and during the Spring Festival, some students even felt sorry for not contributing enough.

“Talk is cheap.” In Yang Li's view, this project has the ability of others to do, but only they go to the ground to practice, and the entire project is completed by first-year students, the process is far more important than the result, they "represent the new force in the field of AI." ”

It hasn't been all smooth sailing with this project.

Jiang Zhengyang, a member of the project and a 21-level artificial intelligence major in the School of Information of China Jiliang University, told Leifeng Network that when the team was modeling, either the network was too large to train too slowly, or the network was too small to meet the requirements and difficult to achieve the expected goal. At the same time, training will also encounter the situation of insufficient computing power.

After many failed attempts, the team had to turn to Professor Yang Li, who added a network structure in which the model became relatively "light" and the training could be expected.

In the end, the team successfully developed the "snowboarding AI refereeing technology". This technology can accurately identify whether athletes grasp the board in complex scenes such as blurry pictures, high-speed camera movements, and long-distance panoramic pictures, so as to provide a basis for referees to score and help "Winter Olympic Fairness".

"Our expertise is limited and we need to continue to strengthen theoretical learning. Through this project, we have learned about the process, methods, and difficulties of doing projects from scratch, and accumulated experience. Of course, in the end, when I saw the results of the project, I was very happy in my heart. Jiang concluded.

Yang Li believes that it is normal to encounter problems, and the key lies in taking action and practicing. "People will fall many times on the road of learning to walk, but they can't just learn to climb because they fall, so they will never go."

This is not a microcosm of the development of domestic AI.

After experiencing the obscure period of Taoguang, domestic AI began to flourish in 10 years, and a number of AI companies such as SenseTime, Megvii, Yuncong, and Yitu were born successively, which were warmly welcomed by capital and supported the hope of domestic AI. However, after the passion burns, it is followed by various doubts such as the difficulty of landing the industry, the difficulty of commercialization, and the difficulty of monetization.

Today's AI is in the groping period from climbing to walking, bumping, falling and falling, etc. occur from time to time, and it is also ridiculed by the public as "artificial mental retardation".

But Yang Li is not frustrated by this, but optimistic, because "there are more and more enterprises, more and more talents to participate in the development, promotion, and landing of AI", under the impetus of the "government, industry, education and research" model, AI will also be unveiled mystery, revealing the most real appearance, and the public will also form a "comprehensive and objective" understanding of AI in the future.

In the process of Leifeng Network's communication with a number of AI practitioners, almost everyone is full of hope for AI, even if AI is still in the "weak artificial intelligence" stage, they still firmly believe that AI has a bright future.

"The sea of AI is not just about the corners and corners, but about changing the world." The engineer who opened the article complaining about "artificial mental retardation" told Leifeng Network that even if the road to changing the world is full of ups and downs, "because of love, so insist." ”

For some ridicule and doubts from the public, he hesitated for a moment and replied:

"Please give AI some tolerance."

For more articles in the field of smart city + AIoT, please pay attention to the public account of Leifeng Network's channel "AI Nuggets".

END