Dialogue: What Learning Tools are Needed for AI Education in Primary and Secondary Schools

In April 2022, Wu Junjie, Dai Juan, and I conducted a session of “chatting” about the development of artificial intelligence learning tools, published in the “Dialogue” column of China Information Technology Education. The core content of the “chat” was to introduce our next work focus: designing a simple introductory tool for AI education in primary and secondary schools, similar to Scratch in programming learning, Arduino in single-chip learning, and control boards in IoT.

Due to the lengthy nature of the “Dialogue,” I did not publish it on the public account in a timely manner. It was only during the May Day holiday that I had the time to organize and publish some of the content. Upon rereading this “Dialogue,” I was also deeply moved: for instance, more than half of the development team has changed, and I myself have left Wenzhou Middle School. Fortunately, the pits we dug are still being filled, and the promises made have basically been fulfilled. Two years later, MMEdu has been upgraded to XEdu, adding tools such as BaseNN, BaseML, BaseDT, and XEduhub, becoming increasingly powerful and gaining recognition from many frontline teachers. XEdu and the PuYu platform have become the best choices for implementing AI education in primary and secondary schools.

I have been engaged in maker education since 2012, and now it has been exactly twelve years. Twelve years has a special meaning in the Chinese context (a full cycle). Sometimes I feel quite emotional; it is indeed difficult for people to break through their inherent cognition. In China, many teachers and enterprises who worked on maker education alongside me back then still shamelessly believe that single-chip microcontrollers combined with sensors constitute AI education. Fortunately, my peers who promoted maker education in China back then, such as Wu Junjie, Zhou Maohua, Guan Xuefeng, and Liang Senshan, have been able to step out of the misconception of “general AI” and bravely stand at the forefront of the tide.

I still clearly remember an afternoon in the fall of 2021 when Dai Juan came to Wenzhou to find me, saying that the laboratory needed to form a team for the Intelligent Education Center to promote AI education in primary and secondary schools. I asked:

Can we develop a quick-start learning tool for primary and secondary school students?

Dai Juan replied:

Sure, I believe in your judgment. But you have to lead the team to develop this learning tool yourself.

XEdu was born from this.

The documentation for XEdu can be found at: https://xedu.readthedocs.io/

Dialogue Guests:

Wu Junjie from Beijing Normal University (a PhD student at Beijing Normal University, initiator of the inclusive curriculum for maker education, and an expert in evaluating AI maker literacy capabilities at www.chinaaitest.com. His research focuses on the evaluation of maker education and professional development of teachers, interdisciplinary project courses guided by science and technology education, and training of top innovative talents.)

Dai Juan from Shanghai Artificial Intelligence Laboratory (Director of the Intelligent Education Center at Shanghai Artificial Intelligence Laboratory, Dean of the Education Research Institute at SenseTime Technology, former Product Director at Apple Siri in the USA and Senior Product Manager at Microsoft Windows Phone.)

Xie Zuoru from Wenzhou Middle School in Zhejiang Province (Head of the AI Innovation Center at Wenzhou Science and Technology High School, Senior Teacher, Provincial Special Teacher, Provincial “Ten Thousand Talents Program” Teaching Master, Vice Chairman of the Modern Educational Technology Branch of the China Electronics Society, Executive Director of the Information Technology Education Committee of the China Educational Technology Association, AI innovation education consultant at the Intelligent Education Center of Shanghai Artificial Intelligence Laboratory, Chief Editor of the high school information technology textbook published by Zhejiang Education Publishing House, and Chief Editor of the information technology textbook writing group for Tsinghua University’s “One-stop Teaching Research” experimental textbook.)

Theme 1: The Relationship Between AI Education, Maker Education, and Smart Education

Wu Junjie: In this issue of the “Dialogue” column, we will discuss the topic “What tools are needed for AI education in primary and secondary schools?” Currently, Teacher Xie Zuoru is leading the development of an AI education tool. To be precise, he is dissatisfied with the current AI learning tools and has decided to “come out of the mountains” to create a new one himself. Of course, the main body developing this tool is the algorithm researchers and engineers from the Shanghai Artificial Intelligence Laboratory (hereinafter referred to as the laboratory), so we have also invited Teacher Dai Juan, the head of the Intelligent Education Center of the laboratory.

Before we officially discuss, I would like to ask Teacher Xie Zuoru a question: In the past two years, “AI education” has become a buzzword in primary and secondary schools. Can we say that AI education can replace maker education, or that AI education has already replaced maker education?

Xie Zuoru: I do not agree with the statement that AI education has replaced maker education. AI education and maker education have never been in a relationship of replacement or opposition. Although maker education emphasizes learning through creation, AI education focuses on learning technology and using AI technology to solve problems. However, the two are closely related because makers always use more advanced tools to solve problems. When the threshold for AI is lowered and it becomes “massively amateur,” makers will naturally use AI as a tool to create intelligent objects and solve more problems.

When I teach AI classes, I consciously look for real problems around us as case studies for students to solve. Maker education emphasizes “learning to solve problems,” and the problems that AI education aims to solve should not be limited to the computer itself but should originate from the real physical world to stimulate students’ interest in learning. Therefore, AI education will adopt a project-based learning concept similar to maker education. For example, over the past decade, I have often worked with students in maker spaces on various technologies for recognizing human postures. These research projects can be considered both maker education and AI education. Thus, AI education and maker education are complementary, not opposing.

[Image] Students in the maker space at Wenzhou Middle School researching human posture-controlled games

Wu Junjie: I remember that in 2017, we organized a small seminar in Beijing to discuss how to carry out AI education with AI industry experts. At that time, the State Council had just released the “New Generation Artificial Intelligence Development Plan,” and artificial intelligence became a buzzword, but primary and secondary schools did not know how to implement it. Later, we organized the main points of the seminar into a dialogue, titled “Artificial Intelligence Education: Disenchantment, Progress, and Practice.” Five years have passed in the blink of an eye; have any of our views changed?

Xie Zuoru: Yes, I vividly remember that meeting. We discussed for a whole afternoon and concluded: AI education has at least two educational goals: one is to adapt people to the AI era, and the other is to develop people who can create AI applications. For primary and secondary schools, we need to first address the first goal; every child must learn how to coexist peacefully with AI. The next step is to attract some children to develop an interest in AI and try to use AI to solve some problems. To achieve this, we need a series of AI projects and a set of AI courses, as well as good AI development and learning tools. The conclusions from the 2017 meeting can still guide current AI education in primary and secondary schools.

Wu Junjie: Now there is also an educational term called “smart education,” which seems to have many connections with artificial intelligence education. Can you briefly analyze this?

Xie Zuoru: Because people often simplify artificial intelligence education as “intelligent education,” which looks very similar to “smart education.” In fact, “smart education” is part of educational informatization, focusing on the application of new technologies represented by the Internet of Things, big data, and artificial intelligence in education, which is different from the education we are discussing today that uses artificial intelligence as learning content. However, I have always regarded many application scenarios in “smart education,” such as smart classrooms and smart campuses, as practical fields for students to study artificial intelligence technology, allowing students to transition from users of smart education to engineers and designers.

Theme 2: The Relationship Between AI Courses and Learning Tools

Wu Junjie: Teacher Dai Juan is not only the head of the laboratory’s Intelligent Education Center but also a seasoned researcher in AI education for primary and secondary schools. She began developing AI courses with her team as early as 2018. I would like to ask Teacher Dai Juan to share her insights after years of deep engagement in AI education for primary and secondary schools.

Dai Juan: I am an industry insider in the AI field, having studied AI and worked for many years in AI product departments at Apple and Microsoft. In 2018, a chance opportunity led me to embark on the path of youth AI education. This is a pioneering endeavor, and our ideas and understandings have evolved based on feedback from actual teaching activities. The typical definition of artificial intelligence is to create machines that can think, learn, reason, make decisions, and act. Based on this, we strive to develop courses that allow students to experience and apply AI, helping primary and secondary school students better understand AI technology and its applications.

As our research deepens, we gradually realize that AI education also has another significance for children’s growth. Because AI is also a kind of cognitive science, when we teach children to train computers to understand language and compose texts, they need to deeply understand the nature of language; when we teach children how to teach machines to learn and understand the methods of machine learning, they need to think about learning itself; when we teach children how to teach machines to reason and make decisions (machine thinking), they need to study the nature of thinking. The human brain has the ability to think about itself, and AI, as a form of artificial external intelligence, can help children formalize their mental and cognitive development into tangible learning processes.

Wu Junjie: Teacher Xie, who has long worked on the front lines, can you evaluate the current state of AI education in primary and secondary schools and share your expectations for AI education?

Xie Zuoru: I believe that from 2017 to now, AI education has not developed well. Although there are many courses named AI education in primary and secondary schools, their quality varies. I categorize the current AI courses in primary and secondary schools into three types: the first type is “fishing in troubled waters,” which simply renames previous programming and robotics courses, including LEGO building courses; the second type is “highbrow,” which directly takes university-level AI courses for primary and secondary schools, but the code is too complex and can only be experienced; the third type is “scratching an itch,” or AI application courses. These courses delegate the AI component to hardware, such as smart cameras and offline voice recognition modules. For students, using a more powerful electronic module to create intelligent works does not mean they understand the principles of AI.

In my view, true AI education is the fourth type of course. Through this type of course, students can master machine learning methods and experience the entire process from data organization, model selection, training the model to ultimately solving problems. When students have personally experienced using machine learning to solve problems, they can deeply understand the significance of data, algorithms, and computing power to artificial intelligence. In 2019, I worked on writing AI textbooks for middle schools for Zhejiang Education Publishing House and Tsinghua University Press, hoping to create such courses, but many ideas could not be realized due to the lack of good AI learning tools. For beginners, the existing AI development frameworks are too difficult; not to mention TensorFlow and Pytorch, even Keras has too high a coding requirement for middle school students.

Wu Junjie: I have been a maker teacher at Jing Shan School for over a decade, and I understand the importance of learning tools for course development. I remember that before the emergence of open-source hardware Arduino, learning single-chip microcontrollers was quite difficult; before the advent of control boards and SIoT (an open-source MQTT server), learning IoT was also quite difficult. My understanding, along with Teacher Xie’s, is that we are both looking for good programming learning tools and have found Scratch. Now, I have come to understand the original intention behind the development of AI learning tools in the laboratory—because the current AI development tools are too difficult for students to master, while graphical tools that claim to enable AI programming are too simplistic. So, Teacher Xie, can you share your thoughts on what a good AI learning tool should be like?

Xie Zuoru: Over the past few years, I have been searching for a good AI learning tool. I hope this tool is easy to start with, allowing students to complete deep learning training with simple code, and the trained model can be deployed as a real AI application. This tool should also allow users to experience the latest AI algorithm models. In simple terms, this tool has characteristics of both a learning tool and a development tool, just like Python, which has simple code and can be used to solve real problems.

Wu Junjie: I agree with your analysis. Since we can’t find such a learning tool, we should create one ourselves, right? After all this discussion, I’ve forgotten to ask what the AI learning tool being developed by the laboratory is called.

Xie Zuoru: MMEdu originates from the open-source AI algorithm framework OpenMMLab, optimized for syntax for primary and secondary schools, lowering the technical threshold, hence the name MMEdu.

Theme 3: What Kind of Tool is MMEdu

Wu Junjie: Can you briefly introduce the functions of OpenMMLab and the members of the MMEdu project team? I think these details need to be introduced by Teacher Dai Juan.

Dai Juan: OpenMMLab originated from the Multi Media Lab (MMLab) at The Chinese University of Hong Kong, led by Professor Tang Xiaowu. The Shanghai Artificial Intelligence Laboratory released the next generation of OpenMMLab at the 2021 World Artificial Intelligence Conference, which has become the most influential open-source algorithm platform in the field of computer vision during the deep learning era. After the upgrade, OpenMMLab covers a wider range of algorithm fields and application scenarios, achieving full-chain value from training to deployment, and is used by many domestic and foreign universities and enterprises for academic research and business production.

The starting point of the MMEdu project is to “dimensionality reduce” OpenMMLab, hoping to lower the technical threshold so that primary and secondary school students can use it. Currently, the person in charge of the MMEdu project is Teacher Xie Zuoru, and the development team mainly consists of interns from the laboratory, who are senior students or graduate students majoring in AI from Shanghai Jiao Tong University, the University of Chinese Academy of Sciences, Shanghai University of Science and Technology, and other universities. Most of them are users of OpenMMLab and possess strong AI research and development capabilities.

Wu Junjie: I understand now; we are leveraging the scientific research achievements of the OpenMMLab team, continuing its hard power in industry and academia to develop a genuine AI tool for primary and secondary schools, guided by the principle of “pragmatic governance”.

Dai Juan: Yes, as mentioned earlier, the launch of the MMEdu project is primarily driven by expectations for future AI education. The laboratory is a national-level research institution. According to the State Council’s documents, research institutions must also undertake the work of AI popularization, meaning that letting more people understand AI is also one of our responsibilities. The laboratory is also an open platform, and by providing good AI learning tools and teaching cases, we hope to establish a vibrant community and ecology for youth AI education, attracting more AI researchers, practitioners, teachers, and students to contribute, cultivating a reserve army of talents with AI thinking for China.

Wu Junjie: Earlier, Teacher Xie mentioned that he participated in writing several versions of AI textbooks. Can you discuss the development of MMEdu from the perspective of textbook writing?

Xie Zuoru: Most of the AI tools used in current primary and secondary school textbooks are basically foreign software developed by overseas teams. For example, in the high school textbook “Introduction to Artificial Intelligence,” all six versions of the textbooks that involve machine learning and neural networks use Google’s AI development framework TensorFlow. Since TensorFlow has a front-end tool called Keras, all neural networks in the textbooks are built using Keras. As I mentioned earlier, the use of Keras is relatively high in terms of entry threshold, and it was not originally developed for primary and secondary school learning. However, at that time, we could not find a better tool than it.

It is precisely because of this that the laboratory’s AI scientists hope to contribute their strength to change this situation. However, they are also very clear that the laboratory’s heavyweight tools such as OpenMMLab, OpenDILab, and Shusheng (INTERN) are only suitable for use by researchers in higher education. As for how to lower the technical threshold and to what extent it should be suitable for primary and secondary school teachers and students, this requires the participation of educational experts from primary and secondary schools in the research.

Additionally, as an educational maker and a long-time information technology teacher who focuses on artificial intelligence, I have also been looking for a low-threshold AI development tool that allows students to truly experience the development process of artificial intelligence, from data to models, and then to deploying AI applications to solve problems, which is the complete machine learning process. However, due to the lack of good AI tools, the machine learning cases in the textbooks are limited to solving simple classification problems such as iris recognition and handwritten digit recognition, and rarely involve more complex AI recognition tasks, such as how many watches or pens are in an image.

Wu Junjie: So you believe that most of the examples in AI textbooks are too simple in difficulty and cannot be used to solve real-life problems with this AI knowledge?

Xie Zuoru: I remember a friend once said something quite unkind, believing that the learning content in many current AI textbooks is like learning “dragon-slaying techniques,” which does not stimulate students’ interest in learning. In the process of compiling high school AI textbooks, we can indeed see the problems of single scenarios, narrow applications, and outdated knowledge. I, along with Professor Fan Lei from the curriculum standards group and Teacher Li Qi from Zhuji, have discussed this topic, noting that the vast majority of high school AI textbooks have not kept pace with the times, where machine learning algorithms remain at the level of decision trees and regression, and the coverage of neural networks is very shallow, even avoiding deep learning entirely. The “Introduction to Artificial Intelligence” module is an elective course, and once students choose it, it means they are on a path of in-depth learning of artificial intelligence, but the textbooks do not provide this path, failing to teach students the most critical and useful knowledge, which is unacceptable.

However, we cannot blame the textbook authors for this; without good tools, no one can do anything, and as the saying goes, “a clever housewife cannot cook without rice.” I am no exception; I am not satisfied with the AI textbooks I wrote a few years ago, which is why I have such a strong desire to develop a good tool.

Wu Junjie: I have also taught AI courses and discussed machine learning and ant colony algorithms, but I have always felt that deep learning is quite profound and I have not studied it. So, can middle and primary school students master these techniques that I still find difficult to understand?

Xie Zuoru: The current wave of artificial intelligence has arisen from the success of deep learning. However, the principles behind deep learning still face a technical bottleneck, which is “inexplicability.” That is to say, currently everyone uses deep learning as a black box, which is why AI engineers jokingly refer to the model training process as “alchemy.” For primary and secondary schools, it is not necessary to fully understand the principles (mathematical formulas) before using deep learning technology. For example, now almost every child can use a tablet to surf the internet and play games, but do they know the principles behind “touch screen” technology? Do they know about ICP/IP and HTTP protocols? From this perspective, we should let primary and secondary students use AI as a “black box” first, and the mathematical principles can be studied later in higher education; this sense of mystery can also motivate them to explore the unknown.

Based on this understanding, I particularly want students to be exposed to some cutting-edge AI applications, such as object detection and adversarial neural networks. The first time I saw a case of adversarial neural networks was in a maker space, where a student used Baidu’s PaddlePaddle to run the classic example of automatically generating cat faces, and the classmates surrounding me were just as excited as I was. Unfortunately, the code was too complex, and I could not include such examples while writing the textbook. Therefore, when I learned that OpenMMLab was willing to “dimensionality reduce” for primary and secondary schools, I was particularly excited, which is how the story of MMEdu began.

[Image] Using OpenMMLab’s MMGeneration to achieve image style transfer

Wu Junjie: Aren’t you worried that once this tool is developed, it will also be like those programming languages that use AI open platform interface technology, where students just write a few lines of code to call it without truly learning the technology?

Xie Zuoru: Let’s take a look at two pieces of code.

Code 1

from aip import AipBodyAnalysisimport jsonAPP_ID = '****'API_KEY = '****'SECRET_KEY = '****'client = AipBodyAnalysis(APP_ID, API_KEY,  SECRET_KEY)def get_file_content(filePath):    with  open(filePath, 'rb') as fp:         return fp.read()image = get_file_content('test.png')print(client.gesture(image))

Code 2

from MMEdu import MMClassification as MMClassificationimg = 'test.png'model = cls(backbone='LeNet')model.checkpoint = '../hand_gray/latest.pth'result = model.inference(image=img)model.print_result()

The first piece of code calls an AI open platform, while the second piece is the model inference code of MMEdu. Both serve the same function of recognizing a gesture image and outputting results. If we remove some necessary user information from the open platform, both pieces of code are very concise and easy to understand. However, the difference is that the former calls a public AI model deployed in the cloud, while the latter uses a locally trained model weight file (“latest.pth”) for inference. The former is like using a “translation software” to read English, without truly understanding English, while the latter genuinely masters the technology, where the model is trained by them and can be trained for different AI models according to their needs.

So, is it difficult to train an AI model with MMEdu? Let’s look at a piece of code used to train an image classification model with the MMClassification module.

Now, has Teacher Wu noticed that both the model training and inference codes in MMEdu are very concise and easy to understand, reading almost like pseudocode? In fact, deep learning is not complicated; it is simply about finding a set of data (dataset), building a neural network model, and then starting training. When the trained model achieves good recognition results or finds that further training will not improve accuracy, training can be stopped, and the corresponding weight file can be saved, as shown in the figure.

Basic Process of Deep Learning

Why can MMEdu achieve such simplicity in code? Please take a look at the line “model = cls(backbone= ‘LeNet’)”; “LeNet” is a typical neural network that works particularly well for recognizing handwritten digits and letters, and is essential content in convolutional neural networks. If we were to write code using Keras to build the LeNet network, it would require a lengthy code segment, while MMEdu can directly call the network name. This is the advantage of using OpenMMLab for AI research. I would like to ask Teacher Dai Juan to explain the underlying principles, as she participated in the development of OpenMMLab when she was studying at The Chinese University of Hong Kong.

Dai Juan: When I was a master’s student at The Chinese University of Hong Kong, I was a student of Professor Tang Xiaowu, and I learned in MMLab. The current person in charge of OpenMMLab, Lin Dahua, once summarized it this way—OpenMMLab helps developers shorten the path of AI projects. Why is that? It is because OpenMMLab has many built-in classic networks like “LeNet.” Models such as MobileNet, yolov, etc.

These built-in classic network structures are also called “SOTA models,” SOTA is short for “state-of-the-art,” referring to the best and most advanced models in a certain research task. There is a saying in the tech community: “Don’t reinvent the wheel.” This means there is no need for developers to laboriously rebuild these SOTA models; they can be called directly. Therefore, OpenMMLab has won the favor of many AI researchers.

Xie Zuoru: This function of OpenMMLab reminds me of Python. For example, in sorting, we used to teach VB, so we had to learn sorting algorithms like bubble sort and selection sort. Now with Python, to sort, you can just use “sorted” or “sort()” directly, which is very convenient. Required textbooks no longer introduce sorting algorithms, but only when studying algorithms (such as in the “Algorithm Basics” module) will we write code to demonstrate the principles of sorting. Similarly, MMEdu, while incorporating many SOTA models, also allows users to build personalized models layer by layer through the BaseNN module (another tool of XEdu).

Wu Junjie: I understand now. Indeed, the emergence of open-source hardware is also to reduce the difficulty of developing single-chip microcontrollers. Controlling the pins with Arduino’s pin code is very simple; for example, the code for PWM and servo control is directly encapsulated as functions, eliminating the need to write them individually. In this regard, the experience of maker education provides excellent insight for Teacher Xie in developing AI learning tools. So, what other work has MMEdu done in simplifying code?

Xie Zuoru: In addition to incorporating SOTA models, MMEdu has also standardized the format of datasets, such as using ImageNet for image classification and COCO for object detection. Students organize their data according to the dataset requirements and specify the path in the code. We believe that organizing datasets is also a universal skill, akin to physics experimental literacy, which can be termed AI experimental literacy.

Additionally, MMEdu simplifies various parameters during training. Since OpenMMLab was initially designed for researchers, it has very powerful features and supports many parameters during training, such as learning rate (lr_config), iterations, optimizer, etc., totaling over twenty. We asked the project team to set default values for these parameters. This way, while sacrificing some performance, the entry-level experience becomes simpler. Currently, training with MMEdu generally requires writing only one parameter, which is Epoch, the number of training rounds.

It should be noted that while all training parameters in MMEdu have default values set, if students understand the value of a particular parameter, they can add it during training. For example, we default “validate” to “True,” meaning the accuracy can be viewed during training; changing it to “False” can improve speed slightly. This suggests to students that the more AI knowledge they master, the better the trained model may be, and the faster it may run. The way these parameters are added is completely consistent with OpenMMLab, meaning that when students master MMEdu, they can seamlessly transition to OpenMMLab.

Wu Junjie: Your statement has also inspired me. In the field of physics curriculum and teaching theory, there is a term called concept progression, which refers to students’ understanding of concepts resembling steps. For instance, the concept of “force” is an intuitive feeling from the body in elementary school. In middle school, “force” can be measured using tools like force meters, showing specific values. In high school, it progresses to Newton’s second law and third law. In university, the concept of “force” may no longer exist, becoming a form of momentum flow. I feel that learning AI through MMEdu is also a progressive process. Now, can we define MMEdu in a short sentence?

Xie Zuoru: My definition is an “out-of-the-box” AI learning or development tool. The term “out-of-the-box” not only refers to built-in SOTA models but also includes the configuration of the environment. Currently, MMEdu can be used directly after unpacking (XEdu provides a one-click installation package), so teachers no longer have to worry about environment configuration.

Wu Junjie: I completely agree. If it can be used right after unpacking, frontline teachers will benefit greatly. Let me ask a basic question. Generally, teachers believe that calling an AI open platform can achieve many AI applications and feel that is sufficient, thus lacking motivation to learn further. So, what is the difference between students using MMEdu to train models and using AI open platforms? Can you provide a specific example?

Xie Zuoru: Indeed, current AI open platforms are very powerful and can achieve many AI recognition functions. For example, in a certain AI open platform, we can upload a photo of an animal and receive highly accurate recognition results. However, if users want to identify their own pets, they need to enable the “EasyDL Custom Animal Recognition” feature, which, although it does not require coding, is a very cumbersome process and incurs fees. On MMEdu, students just need to take several photos of their pets from different angles (this work needs to be done regardless of the tool) and then find some other animal photos, dividing them into two groups, and start training with the previous code. As long as the dataset is well organized, the recognition results will be quite good.

In simple terms, AI open platforms provide general AI capabilities, while self-trained models achieve personalized AI capabilities. Most importantly, the AI problem-solving ability gained through training models with MMEdu is real and permanent, unlike the former, where once you leave the AI open platform, you lose the AI recognition ability. Using AI open platforms to solve problems is actually not directly related to artificial intelligence education; it still belongs to programming education. Artificial intelligence focuses on how intelligence is generated, namely how models are built and trained.

Wu Junjie: This is a persuasive example, and I am intrigued. I remember Teacher Xie mentioning many tools earlier, such as TensorFlow, Pytorch, and Keras, as well as PaddlePaddle. What is the relationship between MMEdu and these tools?

Dai Juan: I can explain this. TensorFlow, Pytorch, and PaddlePaddle can be considered AI development frameworks, while Keras and OpenMMLab can be referred to as front-end API tools that are encapsulated based on AI development frameworks. MMEdu is a further encapsulation of OpenMMLab. Currently, TensorFlow and Pytorch are the two largest camps in the AI field, while PaddlePaddle has a smaller user base. In fact, Keras also supports Pytorch, and OpenMMLab can be modified to support TensorFlow or PaddlePaddle.

Wu Junjie: Can I ask, can the models trained with MMEdu run on mini computers like Raspberry Pi and Vigu?

Dai Juan: Currently, MMEdu can already be deployed on Raspberry Pi and Jetson Nano. However, do not expect to train models on mini computers, as the speed is really too slow. Additionally, there is a module in OpenMMLab called MMDeploy, which is a model deployment tool designed to deploy the algorithm models generated by machine learning training to various mobile or edge computing devices, enabling efficient operation and applying algorithm models to various tasks in real life. Currently, MMDeploy supports algorithm modules for detection (MMDetection), segmentation (MMSegmentation), classification (MMClassification), editing (MMEditing), and text recognition (MMOCR), supporting backend inference engine types such as ONNX Runtime, TensorRT, and OpenVINO. I believe that in the near future, models trained with MMEdu will be able to be deployed on more mini terminals (Note: MMEdu models support direct conversion to ONNX format, which can then be deployed on most intelligent terminals using XEduHub).

Xie Zuoru: Additionally, MMEdu also includes libraries for interaction with open-source hardware or IoT, such as siot, pinpong, and PyWebIO, allowing students to train models on their own computers or computing servers and then deploy them to various open-source hardware. In the AI laboratory at Wenzhou Middle School, several AI projects designed by students that interact with smart home devices have already been deployed. Students have truly regarded the smart campus as a practical field for AI technology.

Theme 4: The R&D Process of MMEdu

Wu Junjie: Although I did not participate in the R&D of MMEdu, I can imagine that this is a pioneering endeavor and must have been challenging. Can you tell me about the difficulties encountered during the R&D process of MMEdu and how they were resolved?

Xie Zuoru: MMEdu started planning in November 2021 and officially launched in March. During the R&D process, we encountered many difficulties, the biggest of which came from computing power. Because MMEdu is aimed at primary and secondary schools, but deep learning has high requirements for computing power, training in a CPU environment is particularly slow, and some larger models can take more than ten hours to train. Many modules in OpenMMLab are designed for GPU environments by default. I emphasized that the GPUs referred to here specifically refer to NVIDIA graphics cards. I surveyed and found that almost no school around me has GPU equipment in their computer labs. So what should we do? We had to lower the requirements, inventory all the modules in OpenMMLab that support CPU training, test them, and finally focus on image classification and object detection as core modules, creating a one-click installation package for primary and secondary school use.

As for other modules, such as generative adversarial networks and semantic segmentation, which I particularly like, how do we handle them? We are preparing to release a version based on container technology (Docker), which can deploy MMEdu on various computing servers (Note: This work has been completed, and the PuYu, Mo, and OpenHydra platforms have all integrated XEdu). Schools that can afford to purchase computing servers can deploy this container on the server to enjoy the full functionality of MMEdu with GPU versions.

Of course, the lack of computing power also affects classroom teaching, as training with CPUs is simply too slow. Therefore, we focused on data sets, trying to streamline them and provide pre-trained models, allowing teachers to continue training based on pre-trained models during class, which significantly speeds up the process. We also need to devise a mechanism to protect the pre-trained model from being overwritten… it’s really hard to put into words.

Wu Junjie: I can empathize; I also know how important computing power is for AI. I hope the hardware conditions in primary and secondary schools improve soon. So, what are the current R&D plans for MMEdu? I know that MMEdu has released a beta version 0.5; when will new versions and the official version be released?

Dai Juan: The official version of MMEdu is planned to be released at this year’s World Artificial Intelligence Conference (Note: The official release of MMEdu was in October 2022). Before that, multiple test versions will be released, and AI teachers from all over the country will be invited to evaluate them. We have also carefully designed a testing manual, which is essentially a simple deep learning tutorial; participating in the testing is equivalent to attending a deep learning training.

Theme 5: Future Prospects of MMEdu

Wu Junjie: After discussing so much, I am becoming more and more excited. Let us look into the future; if there were a ranking of AI learning tools for primary and secondary schools worldwide, what ranking would you hope MMEdu to achieve?

Xie Zuoru: I hope MMEdu can rank among the top, becoming one of the most commonly used and excellent AI learning tools. The reasons are simple: first, it originates from the best AI tool, OpenMMLab; second, MMEdu is backed by a national-level AI research institution; third, this tool is developed by frontline teachers and AI industry experts together, with many teachers contributing their wisdom to this tool, such as Liang Senshan from the Ministry of Education, Yu Fangjun from Shandong, Chen Xiangqun from Shenzhen, and Zhang Jingyun from Jiangsu, as well as other makers from enterprises like DFRobot’s Qiao Yingjie.

When the official version of MMEdu is released, we will also provide an English version of the documentation. Why provide an English version? Because MMEdu is not designed solely for Chinese children; it is designed for all children, just like Arduino and control boards, which belong to the global education commons.

Wu Junjie: Returning to a traditional question, OpenMMLab is an open-source tool. Do you require others to open source their training results (referring to models and weight files) using MMEdu? Or can MMEdu be used for commercial projects?

Dai Juan: We indeed hope that more people will contribute their datasets and trained model weights, so we are planning to create a community. The Intelligent Education Center currently has a team dedicated to building this community, but the question is what kind of mechanism can attract teachers and students to participate in this community and open source their research results? Forcing open source through some agreement seems impractical. The laboratory is a research institution, not a profit-making unit. I believe that as long as we persist in doing a good job with MMEdu, an open-source mechanism will gradually be designed, and I invite Teacher Wu to join us in this effort.

Xie Zuoru: MMEdu can definitely be used for commercial projects, just as many makers use open-source hardware to create product prototypes. I particularly look forward to teams using MMEdu for commercial projects, as this would further demonstrate that MMEdu is a genuine AI development tool, not a mere decorative pillow, right?

Wu Junjie: That’s true. I wish the MMEdu project success, and I will introduce this tool to more friends, continuously providing feedback on my insights and suggestions during the usage process.

Dai Juan, Xie Zuoru: Thank you, Teacher Wu, for your support. Let’s work hard together.

Postscript by Wu Junjie

Completing this dialogue has allowed me to recognize a familiar yet unfamiliar Xie Zuoru. I have known Teacher Xie for many years; he has always been researching cutting-edge technologies, while I have been focusing on the popularization of mature technologies. The emergence of MMEdu in the field of AI education for primary and secondary schools enables artificial intelligence to become a mature technology, serving as a general tool for enhancing educational productivity. The repeated outbreaks of COVID-19 have affected many people’s lives, yet they have not impacted the basic stability of the national economy, thanks to the many machines that have contributed their strength, such as efficient rice transplanting machines compensating for the agricultural time lost due to the unexpected pandemic and modern smart ports compensating for the delays in heavy truck transportation caused by transportation disruptions.

The “MMEdu” tool brought by Teacher Xie and Dai Juan from the Shanghai Artificial Intelligence Laboratory reminds me of the French microbiologist Pasteur, who focused on studying microbiology during the Franco-Prussian War, making significant contributions to France’s economic development after the war. The MMEdu team led by Teacher Xie, building upon the long-term accumulation of OpenMMLab, has developed a tool that allows primary and secondary school students to understand artificial intelligence from its core, enabling ordinary people to train artificial intelligence to become a tool for improving their productivity. The hardships and joys in this process are difficult for outsiders to understand but are experiences we hope to share.

Artificial intelligence is science, and science is the wealth of all humanity. Scientists are the most important force for promoting social progress, and scientific popularization is education and also nurturing. True education generates social value directly and makes society better. Through this dialogue, I wish to express my faith in humanity’s power; a new era of “big teaching” in artificial intelligence education is about to arrive.