This article is about 4,000 words and takes 12 minutes to read
Introduction: On August 6, the seminar on the high-quality development of digital intelligence transformation to promote education and the 2023 annual conference of "Experimental Research on Leapfrog Development and Innovation of Basic Education" and "Experimental Research on Leapfrog Development and Innovation of Basic Education" was held in Beijing Kuangou Conference Center, hosted by the Advanced Innovation Center for Future Education of Beijing Normal University, organized by the Institute of Modern Education Technology of Beijing Normal University and the "Mobile Learning" Ministry of Education and China Mobile Joint Laboratory, and co-organized by the New Generation Open Innovation Platform of Smart Education Country.
At the meeting, Lu Yu, associate professor of Beijing Normal University and director of the Artificial Intelligence Laboratory of the Advanced Innovation Center for Future Education, made a report on "Generative Artificial Intelligence and Its Educational Applications", sharing the definition, technical capabilities, application demonstrations, application prospects, limitations and potential risks of generative AI.
Viewpoint | Lu Yu
Organize | 100 people in education informatization
Source | Seminar on the High-quality Development of Education by Digital and Intelligent Transformation
The following content is based on the video sharing of Director Lu Yu
In recent years, artificial intelligence technology has been widely used in many vertical fields, including education, but at the same time, it also has many limitations, such as there are big problems in intelligence and versatility, whether it is the current intelligent evaluation, automatic marking of essays, or personalized learning software platforms, in fact, there are these problems.
And the reason why everyone is talking about generative AI today is actually because it has the potential to solve these current bottlenecks. To take the simplest example, with the blessing of generative artificial intelligence, maybe Apple's Siri or Xiaomi's Xiaoai classmates are no longer as "stupid" as before, and can only do some of the simplest questions to answer.
01
What is generative AI?
Simply put, generative artificial intelligence refers to the use of some pre-trained large models to generate text, images, video, audio and other multimodal content of a technology, it can also be called artificial intelligence generated content, referred to as AIGC.
It generates many modalities, and it is not only in the past two years that it has begun to be used.
In terms of text generation, some people have long begun to do artificial intelligence generation creation of poetry, and now the business field is also doing some marketing copywriting text generation, of course, the most striking is the generation of some interactive chat conversation text, such as ChatGPT.
In terms of image generation, CAFA actually cooperated with Microsoft in the graduation exhibition a few years ago to produce various styles of paintings such as impressionism and abstraction, and the quality was already very good at that time. Now many e-commerce banners and promotional posters are also doing image generation.
In terms of audio generation, at the earliest time, it is possible to do some basic generation of different styles of music, and now it is possible to synthesize lyrics, composition, arrangement, singing, mixing, etc. to generate audio content.
Finally, in terms of video generation, three or four years ago, students used some related technologies to replace avatars and do some content adaptations of movie bridges. Now you may see in today's headlines that many original pictures of the Qing Dynasty and the Republic of China have been restored to color, which is very realistic. These are the concrete results of generative artificial intelligence in video generation.
Generative AI has been around for a long time. In the 50s of the last century, there was the first string quartet "Ilyak Suite" composed by a computer; Before the 10s of this century, there was already the first novel created by artificial intelligence, note that it is a novel, not a simple short essay or composition; By the end of 2022, everyone is most concerned about ChatGPT.
Therefore, generative artificial intelligence itself has a very good historical evolution, and there are many relevant reserves in technology.
Our team has been paying attention to GPT-2 and GPT-3 since about 2019 and 2020, let's not look at the complex technology in it for the time being, only talk about its basic explicit parameter quantity.
GPT-1 has been available since 2018, and in 2000, a few hundred parameters are a large model. But by 2018, GPT-1 parameters reached 1. 1.7 billion. The number of parameters of GPT-2 and GPT-3 is 15 billion and 175 billion, respectively.
ChatGPT is actually an enhanced or improved version of GPT 3.5, and it is not actually a landmark software system version of OpenAI, but it does show great capabilities.
Looking at the number of parameters, it can also be seen that from 100 million to 175 billion, it took only two to three years to show it.
Do you know how large the human brain is? The human brain is composed of biological neurons, and if we consider every synapse in biological neurons as a parameter, the number of parameters in the human brain is about 100 trillion. The amount of parameters of ChatGPT is about 1/1000 of the human brain, but just 1/1000 of the magnitude can already make it do a lot of things, at least no one dares to say that they are more knowledgeable than ChatGPT, right?
So, what kind of capabilities does ChatGPT have? In simple terms, there may be the following four basic competencies:
First, the ability to generate heuristic content. That is to say, it can generate enlightening and creative texts in the course of multiple rounds of dialogue, including poems, stories, reviews, etc., based on some specific theme or given guidance, which is difficult to achieve with previous natural language processing systems.
Second, the ability to understand the dialogue situation. This is also a problem that natural language processing systems have been trying to overcome, that is, the ability to understand context or the ability to understand multiple rounds of dialogue. If you ask Apple's Siri or Xiaomi's Xiaoai, you will find that you can only talk to it for one or two rounds, and after the conversation is long, it has basically forgotten the situation in the entire conversation. But after more than a dozen rounds of conversations with you, ChatGPT can continue to connect some of the specific meanings and situations in previous conversations. This is one aspect of the big breakthrough.
Third, on the basis of having enlightening content generation ability and dialogue context understanding ability, it has good sequence task execution ability. In the process of multiple rounds of dialogue, you can give it some complex tasks to complete step by step.
Fourth, the ability to parse the programming language. ChatGPT can analyze the structure and algorithm of code programs according to the syntax rules, data structures, algorithm construction and programming specifications of a variety of programming languages, and automatically generate code programs or error cause analysis that meet the task requirements according to user task requirements.
02
Generative AI for educational applications
So, what can today's generative AI help teachers do?
I recorded a little video the other day and I wondered, what can GPT-4 help me do if I am a Chinese teacher?
I first entered Zhu Ziqing's "Lotus Pond Moonlight" text, wanting to see if it could help me do some things related to lesson preparation, such as whether I could find the figurative sentences and anthropomorphic sentences in this text, and output them separately.
GPT-4 gives three figurative sentences and three anthropomorphic sentences, respectively, and also indicates the number of pages.
Next, I asked him to find out the poetry and song that Zhu Ziqing cited, output the complete content of the quoted passage, give the source, and summarize the main content.
It gives two quoted poems and songs, one is "Cai Lian Fu" and the other is "Xizhou Qu". What these two articles talk about and where they come from, GPT-4 gives specific information and links to sources, and it summarizes them very well.
Then I thought, can I generate the corresponding pictures from GPT-4 according to the keywords in these three articles, and I put them in the PPT for lesson preparation?
The picture of "Lotus Pond Moonlight" does have lotus pond and moonlight; The picture of "Picking Lotus Fu" is indeed related to lotus; The picture of "Xizhou Qu" does not seem to be so directly related to the article. But, at least we see that GPT-4 can easily generate multiple images corresponding to different ancient or modern texts.
If I'm a new teacher and haven't taken the lesson "Lotus Pond Moonlight" at all, I still want to know, what questions will students ask in class? How should I answer? So I'm trying to get GPT-4 to do some role-playing work, which is something that generative AI is very good at right now.
GPT-4 plays the role of language teacher with teaching experience, students with weak language foundation, and students with good language foundation. After discussing with the Chinese teachers I know, I believe that the GPT-4 answer still meets the basic requirements of teaching.
Next, I asked GPT-4 to generate a multiple-choice question and an essay question, both of which were generated very quickly.
It can be seen that for teachers, especially novice teachers, it is still helpful to ask GPT-4 a few questions when preparing for lessons.
What else do teachers need to do besides lesson preparation? For example, after taking the three classes of "Lotus Pond Moonlight" and getting all the students' essay scores, the teacher is traditionally required to enter the scores into Excel one by one for analysis. But now we can leave it to GPT-4 to do it.
The following video shows the GPT-4 response:
Video loading...
We simply enter a paragraph, and GPT-4 can automatically generate a lot of analysis data, including the student's score band, what proportion of students in different bands, what is the average score of all students, and so on.
It will also automatically select the appropriate diagram, such as a doughnut chart or a fan chart for a basic presentation. In the process, it may go wrong, but it now has the basic code self-healing capabilities, and after fixing, it can continue with the task.
As you can see in the video, it fully lists the percentages of the corresponding bands for Class 1, Class 2, and Class 3. When visualizing, it may feel that it is more appropriate to call the bar chart in this context, so the column chart is automatically generated to show the information of the three classes. At the same time, the main problems were analyzed, and a comparative summary was made between classes.
GPT-4 made another error when generating the radar chart, but it took about 5 seconds to fix the error automatically, and then generated a radar chart with data from three classes.
Is the distinction of this exam good? Is the normal distribution appropriate? Then, GPT-4 made a graph showing the normal distribution, and here, it has actually risen to the point of evaluating essay topics from a teaching and research perspective.
Finally, it analyzes the basic situation and existing problems of the three classes. Of course, it wasn't necessarily right, but with a very brief passage and no new prompts, GPT-4 automatically did this very rich stuff.
This is a simple example that I want to show you.
03
Educational Perspectives of Generative AI
Since GPT-4 has these capabilities, what educational applications of generative AI can we have, and what can we expect?
For generative AI, we generally say that it is based on pre-trained large models. The pre-trained large model itself is a super-large-scale neural network, which is not necessarily possible for the best universities in China now, but on the basis of the large model, we can continue to do its transfer learning.
That is to say, for some specific downstream tasks, you can fine-tune through some small amounts of data, and focus the basic general capabilities of the large model on some specific tasks, such as the generation of teaching materials, teaching resources, teaching evaluation, etc. in education and teaching.
Fine-tuned models can solve education and teaching problems that artificial intelligence has not been able to solve before. In general, in addition to some educational application problems that artificial intelligence has already solved, current generative artificial intelligence can focus on solving the following problems:
First, based on multimodal data and knowledge (whether it is Internet data or subject area knowledge), combined with teaching scenarios, complete some understanding of teaching objects and educational resources. In fact, in high-dimensional space, large models can now make a deeper understanding of these teaching objects and teaching resources. On this basis, it can better understand the teaching objectives and teaching process, and finally make a series of adapted models for educational tasks, including intelligent assistance of teachers, automatic generation of teaching resources, etc., and finally achieve human-computer collaborative process support.
In the example just now, the independent generation of teaching resources such as essay questions and multiple-choice questions is very simple, and the teacher can also adjust the difficulty of these questions and the knowledge points assessed.
There is also teaching intelligence assistance, where generative artificial intelligence can assist teachers to simulate questions and answers between different roles in the classroom. Last month, we did a workshop with Carnegie Mellon University at the world's top AI education conference in Tokyo. We simulated 5 different characters of students to fully simulate the basic evolution of a class. In addition to helping teachers prepare for lessons, many times for students, observing the entire dialogue process through a third-party perspective can actually promote their self-learning.
The combination of teaching resource generation and teaching intelligence assistance can complete the process support of human-machine collaboration.
For example, in the previous example, the teacher will write a teaching summary report in the future, and there is no need to manually generate charts and fill in the text. The teacher can first give the machine a basic task goal, and then keep asking it to adjust the small tasks you want to do, and eventually it can automatically recall the relevant code. Teachers no longer have to deal with code or some programming functions in Office, and machines can automatically generate relevant images and text.
The application of generative artificial intelligence can allow us to do better than before in the three dimensions of automatic generation of teaching resources, intelligent assistance of teacher teaching, and support of human-computer collaborative process.
Generative AI has many benefits, but it does also have many limitations.
First, everyone is talking about it, and it can lead to academic integrity problems. In addition, generative artificial intelligence itself is a large-scale neural network, and the interpretability of general-scale neural networks itself is very poor, and the interpretability of this ultra-large-scale neural network is even worse, so the basis of the content it generates may have many problems.
Its Chinese expression ability is also relatively weak, because whether it is GPT-4 or other versions, it is basically trained on the Latin language corpus, and Chinese only accounts for a small proportion of the corpus, but through the examples we gave, you can see that its understanding of Chinese and language expression, at least at the level of primary and secondary school teaching, is not a big problem, and can generally meet the basic teaching requirements.
In addition, there are intellectual property issues, high costs, etc., which are some of its limitations and potential risks.
Finally, I would like to conclude this brief remarks today with this picture.
When I first looked at this picture, I saw that the tangible hand of generative artificial intelligence put intelligent machines into various collaborative work scenarios in society, and education is also one of these scenarios. But on the other hand, this invisible hand may gradually replace those teachers who are not very willing or good at using artificial intelligence, and strip them out of the team of modern social development, so I also hope that teachers will pay more attention to some application results in this field, try more, and keep up with the pace of development in the era of artificial intelligence.
Note: Lu Yu is an associate professor and doctoral supervisor of Beijing Normal University, and the director of the Artificial Intelligence Laboratory of the Advanced Innovation Center for Future Education. This article is compiled from the "Digital Intelligence Empowerment Education, Classroom Integration and Innovation, Teacher Professional Development" Digital Intelligence Transformation to Promote the High-quality Development of Education Seminar and the National "Basic Education Leapfrog Development Innovation Experimental Research" 2023 Annual Conference.
"Education Informatization 100 People" is a think tank media jointly initiated by industry, education and research media, focusing on education informatization, education digitalization, smart education, Internet + education, artificial intelligence education, education technology and other fields, we want to "make high-quality information and knowledge be seen faster"!