Cui Wei, Chief Scientist of Squirrel AI Learning: Using AI to Scale Personalized Learning | CCF-GAIR 2018

Leiphone.com: The 2018 Global Artificial Intelligence and Robotics Summit (CCF-GAIR) was held in Shenzhen. The summit was hosted by the China Computer Society (CCF), co-organized by Leiphone.com (public account: Leiphone.com) and the Chinese University of Hong Kong, Shenzhen, and received strong guidance from the Bao'an District Government. It is a top exchange event in the three major fields of domestic artificial intelligence and robotics academia, industry and investment, aiming to create the most powerful cross-border exchange and cooperation platform in the field of domestic artificial intelligence.

CCF-GAIR 2018 continues the "top" lineup of the previous two sessions, providing a rich platform with 1 main venue and 11 special sessions (bionic robots, robot industry applications, computer vision, intelligent security, financial technology, intelligent driving, NLP, AI+, AI chips, IoT, investors), intending to present more forward-looking and practical conference content and on-site experience to participants from three fields from multiple dimensions of industry, academia and research.

At the "AI+ Special Session" on the third day of the conference, Dr. Cui Wei, Chief Scientist of Squirrel AI Learning, was the third guest speaker. He delivered a wonderful speech entitled "Let Every Child Have an AI Super Teacher", painting a beautiful picture of future education for us.

Dr. Cui Wei introduced that each student’s learning ability and knowledge mastery level are different, and traditional classroom unified teaching cannot meet students’ personalized learning needs.

Different students have different preferences for teaching styles. Students with good grades may like older teachers who can get to the point, while students with poorer grades may like younger teachers with a relaxed and lively style.

In the past, students relied mainly on the “sea of ​​questions” tactic to make up for their knowledge gaps. In fact, 80% of this was ineffective work because they had already mastered 80% of the questions.

These have seriously affected students' learning efficiency and effectiveness.

To this end, Yixue Education began to develop the artificial intelligence adaptive learning system "Squirrel AI" in 2014. The Squirrel AI intelligent adaptive learning engine includes four modules: student portrait, content model, teaching model and prediction model. The system contains different types of teaching resources, and can accurately match teaching content according to student portraits. During the learning process, the system will monitor and evaluate students' learning effects in real time and diagnose knowledge weaknesses. At the same time, the system can also predict the effect of students learning a certain knowledge point and the time required, so as to formulate the best strategic learning path.

Since 2017, Squirrel AI Learning has held "Man vs. Machine" competitions in many cities, allowing the adaptive learning system to compete with experienced public school teachers in a fair and just manner. The results show that the teaching effect of the artificial intelligence adaptive learning system is better than that of human teachers.

Dr. Cui Wei said that Squirrel AI Learning hopes that AI teachers can solve the problem of learning efficiency, freeing human teachers from preparing lessons, teaching, and grading homework, so that they can engage in more "education" work and cultivate students' imagination and creativity. He believes that future education will be "human-machine co-teaching", and AI super teachers will be a combination of Socrates, Leonardo da Vinci, and Einstein.

The following is the full content of Dr. Cui Wei’s speech. Leifeng.com has sorted and edited it without changing the original meaning:

Good afternoon, everyone. My name is Cui Wei, and I am from Shanghai Yixue Education. Thank you very much to Leifeng.com for providing us with the opportunity to share our application and practical experience in the vertical industry of artificial intelligence + education over the past three or four years as representatives of educational enterprises. The theme of my sharing today is "Let every child have an AI super teacher around them."

Artificial intelligence is everywhere and involves every aspect of our lives. The applications we see are mainly concentrated in the fields of computer vision (including image recognition, security, and security inspection), natural language processing, and text understanding. Image recognition and natural language processing are also used in the education industry, such as some tools for grading homework.

The AI ​​education applications we are working on are somewhat different from these. Our approach is similar to AlphaGo and IBM Watson, hoping that AI can simulate the knowledge and teaching experience of an excellent teacher. Our mission is to teach students in accordance with their aptitude through AI and big data technology, to scale personalized learning, and to enable every child to learn according to their own characteristics. We have been working towards this direction in the past.

Why do we do this? Because traditional classroom education cannot meet the needs of all students. This model is more suitable for teaching new courses, because all students do not know much about the content of the new course and are at the same starting point. However, each student's learning, acceptance and intelligence level are different. After a class, the learning effects of different students vary greatly. After-class review and other factors will also affect the learning effect. As time goes by, the students' mastery of knowledge will gradually widen the gap.

An excellent teacher can usually do two things: first, accurately diagnose students' problems; second, develop personalized learning plans for students. Such an excellent teacher needs decades of teaching experience, which can only be understood but not expressed in words, and is difficult to copy and pass on. Even if it is passed on to new teachers, the latter need to constantly polish and accumulate in teaching practice, just like old Chinese medicine practitioners.

What is particularly important is that the number of students is very large, while the number of teachers is relatively small. Teachers have limited time and energy and it is impossible to take care of every student and provide them with personalized learning.

Artificial intelligence has made great progress in recent years. We believe that artificial intelligence is an extension of human intelligence. Take AlphaGo as an example. Six or seven years ago, people believed that playing Go required a global perspective of humans, which was something that artificial intelligence could not simulate, so it could not replace humans. But AlphaGo overturned this concept. In just a few years, it learned hundreds of thousands of moves in human history and evolved hundreds of thousands of moves. It proves that artificial intelligence can indeed surpass humans in some aspects and do things that humans cannot do. Another example is IBM's Watson, which can imitate excellent doctors to make diagnoses and give treatment plans, and the treatment plans it gives are better than those of ordinary doctors. We believe that artificial intelligence can also surpass humans in the field of education, just like in the fields of medicine, Go, Texas Hold'em, etc., and popularize personalized learning that cannot be scaled under the traditional education model.

When we go to the hospital for treatment, sneezing may be caused by catching a cold, or by fever and allergic rhinitis. Allergic rhinitis has several causes, which may be pollen allergy or dust allergy. Only by accurately finding the cause can we better treat it. The same is true for learning. Only by accurately diagnosing the characteristics and weak knowledge links of each student can we learn more efficiently.

Adaptive learning is a new concept in foreign countries. It has a history of more than 20 years, from the earliest rule-based adaptation to the current artificial intelligence-based adaptation. Adaptive learning can more accurately identify the characteristics of each student and provide them with efficient personalized learning solutions.

In the past, we mainly judged students' mastery of knowledge based on scores. Traditional exams are only a rough assessment of students' knowledge level, which is affected by the scope of the exam (it is impossible for the exam to cover all knowledge points) and the students' test mentality. After adopting artificial intelligence algorithms, we can more accurately locate students' proficiency in each word, comprehensively cover students' mastery of all knowledge points, and timely and effectively allow them to carry out personalized learning for weak points, and learn where they can't. We will accurately recommend the most suitable learning content to students. If the students have a low learning level, we will recommend easy content, and if the students are okay, we will recommend more difficult content. In the past, students usually used the "sea of ​​questions" to make up for knowledge gaps. In fact, 80% of them were ineffective work, because they had already mastered 80% of the questions. Moreover, different students know and don't know different things. In the traditional model, all students' learning content and homework are exactly the same, and it is impossible to teach students in accordance with their aptitude.

Adaptive learning has been applied abroad, covering hundreds of subjects in different countries and age groups, from primary school, junior high school, high school to vocational education. Its effect has been well verified, and it can improve the grades of both primary school students and junior high school students, liberal arts students and science students.

Yixue Education started developing AI adaptive learning systems in 2014, and officially launched the country's first AI adaptive learning product "Squirrel AI" for K12 students at the end of 2015 and the beginning of 2016. Our products completely cover all learning processes. Students only need a computer, an account and a network to log in to the system, and AI will provide students with personalized learning like a teacher.

Here are some examples to show the results of students’ systematic learning. For example, there was a student who didn’t like writing essays at all before, no matter how we asked or encouraged him. After studying with us for a week, he was able to write a 500-word essay.

Another junior high school student used our system when preparing for the high school entrance examination in his third year of junior high school, and his total score increased by nearly 100 points. Our product is particularly effective for reviewing and preparing for the high school entrance examination, and can diagnose all the knowledge gaps of students and help them solve their knowledge weaknesses in a targeted manner.

There was another student who had attended many tutoring classes and used various products, but his English scores had not improved much. After using our product, he improved his score from 63 to 120 through self-study without any teacher intervention. In the end, he was praised by the vice-principal of the whole school. At that time, the school only praised two students, one was a top student, and the other was him.

How do we do this? We monitor and evaluate students in real time during their learning process, build student profiles through data analysis, and understand their proficiency in each knowledge point. Based on this, we then plan the best learning path and learning solution for them.

We can be like an excellent teacher, not only diagnosing and planning for students, but also formulating strategic learning strategies for them. For example, a student facing the sprint for the high school entrance examination has to learn 100 knowledge points within two months. Based on his past learning situation, we judge that he cannot fully master these 100 knowledge points within two months. Then we will give priority to letting him learn the knowledge points that appear frequently in the exam, account for a high proportion of scores, and are easier to learn. Because if all 100 knowledge points are learned, you can only skim the surface, and in the end, you will not master any knowledge point well. If you focus your time and energy on important knowledge points and study them thoroughly, you can score steadily in the exam. This is the most efficient way to learn.

Our AI teacher can also make forward-looking judgments. If a student wants to achieve a certain learning goal, we will first determine whether he is currently capable of achieving this goal. For example, if a student is learning quadratic equations, we will see whether he has learned the decomposition and linear equations well. If not, he will first learn these weak basic knowledge points. Only by mastering these prerequisite knowledge points can he achieve his learning goals.

Our system will recommend the most suitable learning content to students. Different students have different learning styles and interests. Students who are particularly good at learning may prefer older teachers because they can get to the point and students are more efficient. Other students may like relaxed and lively courses. Our system has many different types of learning content, and we will make corresponding recommendations based on the characteristics of the students.

In order to evaluate the learning effect, we have conducted more than ten comparative experiments in a fair and just manner since October 2017. We provide a system, and the local education bureau or public school finds a group of students and randomly divides them into two groups to ensure that the scores of the two groups are consistent. One group is the experimental group, which uses our system to learn independently without teacher intervention; the other group is the control group, which is taught by a public school teacher with more than ten years of teaching experience in the traditional way, with one teacher teaching 15-20 students. We let the teachers make the test papers according to the given learning scope, and the students are tested on one set before learning and another set after learning to ensure the same difficulty.

Since 2017, we have conducted more than ten comparative experiments in Zhengzhou, Nanchang, Wuhan, Jiaxing and other places, and the results have proved that our product is more effective than experienced teachers. The most recent experiment was held in Dongying, Shandong, where we gathered more than 120 students. This comparative experiment was more challenging, with one teacher only teaching three students, and the results proved that our product was as effective as one teacher teaching three students.

Our system includes four major functions: diagnosis before learning, diagnosis during the learning process, recommendation of personalized learning plans and personalized learning content, and planning of personalized learning paths.

Let's talk about the diagnosis before learning. Every student who comes to us must first go through an intelligent diagnostic test. We can quickly and accurately diagnose his mastery of each knowledge point with a small number of questions and a small amount of time. After the diagnosis, there is a report that tells the student which knowledge points he has mastered and which he has not mastered. Next, the student only needs to learn the knowledge points that he has not mastered.

We can also constantly evaluate students' knowledge mastery during the learning process, just like excellent teachers. Just like when a patient is hospitalized, nurses need to frequently measure the patient's blood pressure and body temperature and adjust the treatment plan. The system can timely understand whether the student's ability has improved after completing the quadratic equation problem. The AI ​​engine will automatically recommend the next learning content, adjust the learning path and learning content in time, and ensure that the current learning content is the most helpful to the student.

Our AI adaptive learning engine consists of four major modules:

1. Student portrait: Collecting student data to form a student portrait is the basis for accurate and personalized recommendations.

2. Content model. Our system contains a large number of learning resources, including explanation videos for each knowledge point, and each knowledge point is matched with test questions of different difficulty and levels. With student portraits and content models, we can build a bridge between students and content and achieve accurate personalized recommendations.

3. Teaching model. Provide personalized solutions to students in a timely manner based on their profile. For example, predict whether students are able to master these knowledge points before they start learning, and which knowledge points will help them the most. This is like AlphaGo's dynamic programming. AlphaGo will push back ten, a hundred, or ten thousand steps for the next move to know which move has the highest probability of winning. The same is true for the learning plan we give students. For example, if a student has not mastered a hundred knowledge points, we will push back a hundred steps to determine which knowledge points will help the student the most, and then recommend them to him first.

4. Prediction module: We can also predict what level students can reach on knowledge points that they have not yet learned and how long it will take to learn, so that the AI ​​teacher can make an overall learning plan for students.

The figure below shows the traditional learning path, where all students follow a unified path.

With an AI teacher, there is no need to practice questions like this. For example, if we diagnose that a student has five knowledge points to learn, the system will plan the best learning path.

Another student has more weak knowledge points, and after system diagnosis, it will provide another set of optimal learning paths.

The picture below shows the knowledge gaps diagnosed by the AI ​​teacher for the third student and the corresponding learning path planned.

In our system, each student learns according to his or her own learning path. The AI ​​teacher diagnoses each student's knowledge gaps and assigns him or her different study time, learning content and test questions, allowing students to learn in a personalized and efficient manner.

The figure below is a distribution chart of the time ten students spend on different knowledge points.

The following figure shows the different learning paths of nine students on eleven knowledge points. The first student has only three weak knowledge points, the middle one has seven, and the last one has ten.

Their learning paths are different, and different students spend different amounts of time learning the same knowledge point and have different levels of difficulty in solving the questions.

Our product is student-centered. Students learn independently in the system, and teachers only play the role of answering questions and solving doubts. We hope that AI teachers can help students improve their learning efficiency and liberate teachers. In the past, teachers were busy preparing lessons, teaching, and grading homework, and had no time to educate students. With AI teachers solving the problem of learning efficiency, human teachers can do more meaningful and valuable things, such as emotional communication and exchanges with students, helping students solve problems with learning motivation, and cultivating their creativity and imagination. Artificial intelligence replaces not only simple and repetitive jobs, but also many intelligence-related jobs, including financial high-frequency traders and lawyers, so the cultivation of creativity and imagination is more important.

To improve ability, it is definitely not enough to rely on traditional teachers to teach. The future is the era of "human-machine co-teaching". Learning is completed through the system, and creativity and imagination are cultivated by teachers. We hope that the AI ​​teacher in the future will be a combination of Socrates, Leonardo da Vinci and Einstein.

Exclusive interview with Oxford professor Ivan Martinovic: Biometric-based identity authentication is also not safe. What other black technologies are needed for financial supervision?

Financial technology ignites the GAIR conference, industry, academia and research discuss the path of innovative development | CCF-GAIR