The series of "Punching Smart China" has been updated for several issues. Some readers said that they love to read this kind of down-to-earth real stories. Some readers also reported that they are not electricians, but clerks, farmers, and sand control personnel. Isn't artificial intelligence a high-end subject? ? Where have all those highly educated PhDs gone?
The answer is: they are in the fields, factories, and mines.
Dr. Lin from the Chinese Academy of Sciences, I met him twice.
The first time was in 2020, the opening ceremony of a certain AI elite training class. At that time, Dr. Lin's understanding of AI was still at "the fur of deep learning". He called the platform's machine vision model and developed a model for recognizing animals. He said: "There are still relatively few species that can be recognized. I want to see what other students are doing with AI, so I can learn more from you." When Dr. Lin introduced this AI product, he was still a little unsure.
The second time is in 2022. In an interview in an ordinary conference room, Dr. Lin's speech became more comfortable and richer.
"The previous biometrics application has produced many models in recent years, extending to the identification of mammals, amphibians, insects, and butterfly species," he said, "We also use AI to do Other things, one is the image collection and automatic recognition of wild creatures, and the other is to protect wild animals through voiceprint recognition. The equipment has been completed. We also support other companies to do AI commercial applications, such as agricultural pests and diseases identification".
In two years, Dr. Lin has grown from an AI novice to an AI proficient, and from a scholar sitting in front of a computer doing research, he has gained a lot of experience in going to the deep forest and going to the farmland. The AI in his eyes reflects the upper limit and depth of Smart China.
Today's story is about how a doctor in the scientific research field uses AI and makes AI from scratch.
What's different about the AI in the research institute?
The intelligence in the field of scientific research has both similarities and differences with industries such as industry, agriculture and commerce.
The same points are: 1. Lack of people; 2. Lack of numbers; 3. Lack of scenes.
The field of scientific research has gathered a large number of highly educated talents, but there is still a lack of compound talents who can combine AI with sub-disciplines. Take Dr. Lin as an example. He majored in informatics. Because biological research requires information technology such as computer analysis, he joined the Institute of Zoology under the Chinese Academy of Sciences. He has deep learning, image recognition, voiceprint recognition and other artificial intelligence For technology, he also has to learn from scratch and iterate his own knowledge system.
In addition, most research institutions have a certain amount of data accumulation, but they are also divided into fields, specialties and even species. For example, Dr. Lin wants to do AI applications such as species recognition and bird voiceprint recognition. For many rare animals, not to mention voice data, there are not even a few images of wild activities in the database. Without sufficient data, AI is difficult. Play a role.
In addition, which topics and directions of scientific research can use AI, this is also something that requires scientists to use their "subjective initiative" to explore. Dr. Lin mentioned that there are already very mature computing tools in many scientific researches, and the use of AI must be promoted from the front-line scientific researchers, which requires a certain idea. He said: "So I heard that there are For this in-depth learning talent training, I took the initiative to sign up with my project, hoping to graduate smoothly."
Of course, AI+ scientific research also has its special features.
For example, especially without money.
When we met for the first time, I asked curiously, "Are there many collisions between AI and living things?" Dr. Lin said with a smile:
"AI is being applied very quickly in the medical field, but it is not so good in biometrics. It may have little to do with the commercial value of basic research such as animal protection, and there is a certain lag in investment."
Just kidding, the combination of AI and scientific research is also very important.
Dr. Lin's experience proves that once AI innovation in the field of basic research grows, it can bring about a series of linkage effects in industry, education, research and application.
In 2020, Dr. Lin has just made a popular science app for identifying animals and plants, which integrates computer vision technology and his unit's animal and plant encyclopedia data, which is convenient for animal lovers, field observation investigators, elementary school students and parents to use.
China has a vast territory, and experts have a certain timeline for inspections, such as once every five years. It is very likely that opportunities to observe certain species will be missed, resulting in biased data. Moreover, taxonomists and talents in the biological field are also shrinking, and fewer and fewer people are willing to engage in such arduous field work, so it is very unrealistic to rely solely on researchers and field workers to collect and classify.
Through the innovative application of intelligent technology, more data partners are gathered, and organizations and users are encouraged to participate together, which is very helpful for the development of biological protection and research.
And such a pure public welfare work has subsequently attracted the attention of many research institutions, public welfare organizations, and enterprises/industry circles.
Dr. Lin told me that after the app was developed, many similar international animal protection organizations established good relationships with them. They exported technology, and the other party fed back the biological data obtained to them, allowing them to earn hundreds of thousands of dollars. Many of the image data are newly added observation points, which is equivalent to saving a lot of scientific research funds for the country.
A sustainable data acquisition mechanism is very important for the development of subsequent scientific research tasks.
In addition, they have also attracted some commercial project parties to seek cooperation.
For example, some companies know that they have this technology, and they find out that they want to develop a pest identification system. It happens that the unit where Dr. Lin works is also responsible for the prevention and control of diseases and insect pests, and has the function of ensuring food safety. So the two sides hit it off and started to build a database of pests and diseases and develop identification algorithms.
Compared with industry, agriculture and commerce, which are closer to the economic value, the intelligence in the field of scientific research does not seem to be so urgent, but it has the role of an "innovation engine": it does not affect a person or a company, but may be an industry (such as the above mentioned agriculture), a group (such as popular science education) or even a country (such as discipline talent construction).
From ivory towers to farmland: the footsteps of a Ph.D.
The electric workers and water plant employees in previous reports are all part of the end-link of AI applications.
With the help of Dr. Lin, we can see how an AI capability is born and how it reaches the application end.
The first step: make data from scratch.
After receiving the company's request to identify pests and diseases, Dr. Lin found that data is a big challenge.
From a biological point of view, there are many kinds of pests. There are many forms in the development process of a pest. The developmental states of eggs, weak insects, larvae, and adults are all different, and they belong to different categories in the process of image classification. If It is a layman who does not understand it at all and does not have a formed database.
Therefore, Dr. Lin and his team had to go from the computer desk to the field to collect images of different stages in the field to determine the shape characteristics of different developmental stages of pests and diseases, laying the foundation for the accurate identification of subsequent models.
Step 2: In-depth communication with agricultural experts.
Pests and diseases in different regions may be different, which involves a lot of very professional knowledge points, such as what are the common pests and diseases of tomatoes in the suburbs of Beijing? What kind of control measures will be taken after pests are found? What medications are provided? This requires going deep into the greenhouse and communicating with local plant protection experts and farmers who have been engaged in front-line production for a long time, so as to ensure that the AI system built later can provide a full set of work from pest identification to prevention and control guidance.
Step 3: Model development and tuning iterations.
Dr. Lin and his colleagues are basically informatics majors, not deep learning algorithm engineers, and the model iteration of agricultural projects is very fast. After a batch of images are collected, a batch of models will be iterated immediately, in order to accurately identify different crops , but also to develop different models, so there are many models, and the workload of development and iteration is heavy.
Dr. Lin said that in scientific research work, the flexibility, scalability, and advancement of the AI development framework may be very important. For example, in order to publish papers for peers to reproduce, people are more inclined to use overseas platforms such as TensorFlow and PyTorch. When developing industrial models, we pay more attention to the ease of use of the platform, whether the application-oriented industrialization capabilities are easy to use, and whether there are already trained models with good performance.
Because the underlying technical means, theoretical methods, and domestic and foreign development platforms are the same, and they are all open source, there is no difference, but it is easy to train, easy to deploy, and whether it can be industrially mass-produced (models). The answer to the question is very important for industrial intelligence.
At present, the agricultural pest and disease application supported by Dr. Lin and his team has landed in Beijing. This is also the first app for identifying pests and diseases in the northern region.
Why do you do this project? The starting point of Dr. Lin and his team is different from that of commercial companies - on the one hand, there is a need for enterprises/farmers; on the other hand, it is to help new farmers.
"Now a large number of new agricultural producers, such as the younger generation," Dr. Lin mentioned, "they do not have skilled production experience, nor do they rely on the experience imparted by the previous generation, but rely more on information obtained from the Internet. At this time we Provide him with a tool - AI recognition, and slowly they will grow into experienced producers."
In the past, the protagonists of "Punching Smart China", the drones released by electric workers, and the smart systems used by water plant employees every day, there are countless Dr. Lin behind these AI tools.
students? tutor? researcher? Developer?
The identity transformation of an AI person
The first time I saw Dr. Lin, he was still an "AI novice student" with a project waiting to be judged.
The second time I saw him, he was already a technical expert in an AI open source community in China.
Dr. Lin usually also undertakes some tasks such as novice guidance and development experience sharing, and open source his own data and models on the platform, so that developers and student users have the opportunity to practice and practice.
Cultivating more AI talents, this kind of "altruistic" thinking seems to be engraved in this scholar's mind, and he does it naturally without hesitation or thinking.
It is a recognized fact that there is a large demand and a large gap for AI talents, and it has also become a limitation for the further prosperity and development of the artificial intelligence industry.
It is difficult to cultivate AI talents. One is that the technology is new. Compared with traditional disciplines such as science, mathematics, biochemistry, and mature disciplines such as computers, AI itself is an emerging technology that is still being updated and iterated. There is no perfect talent training model. and the curriculum system; second, strong cross-cutting. AI is an applied technology that requires close integration with industry and practical implementation. School education alone cannot meet the training needs of AI compound talents.
The 14th Five-Year Plan proposes that the artificial intelligence industry should form a technological innovation system with deep integration of industry, learning and research, and connect enterprises, universities, research institutes, governments and other innovation subjects.
In the AI ecology, "industry-university-research-application" is layered, and each layer undertakes its own main innovation tasks, but layering does not mean separation, and each layer must be closely linked to form an innovation chain.
"Dr. Lin" has become the key role that wanders through the layers:
In industry, they are developers of AI innovations;
In the cultivation of talents in colleges and universities, they are the seniors who guide the juniors and juniors;
In research, they are pioneers in combining AI technology with scientific exploration;
In the application link, they will take the initiative to get close to the front-line users to make AI tools more useful.
"There are great Confucianists in talking and laughing, and there is no white man in communication." This is today's AI scholars, who not only have the ideals of being in the ivory tower and helping the world, but also have the ability to take the lead and go deep into the industry.
A teacher at a technical university once told me that AI research should be implemented in factories, which is similar to Dr. Lin's implementation of AI in farmland.
The experience of these scholars embracing AI made me deeply feel that AI is both a profession and not a profession. AI will draw more experts and scholars into the world of AI, and the closed loop of "industry-university-research-application" will accelerate to continuously open the upper limit and imagination space of AI.
The movement of Dapeng is not the lightness of a feather. Only the joint efforts of multiple talents can support and lift up a smart China.
