The big model of education welcomes new players: Ten years of hard work points to intelligent adaptation, 25 questions can test 1,000 knowledge points

Bai Jiao is from Ao Fei Temple
Qubit | Official account QbitAI

In the new year, big models in the field of education welcome new players.

From Squirrel AI , a person who has been involved in the education field for ten years, has just released the first large model of intelligent adaptive education.

According to reports, it can support more tasks besides plain text, and has better performance in knowledge acquisition, information input and conveying information. It is essentially different from traditional large language models. According to the plan, this large model will be installed in Squirrel AI’s existing learning machine products.

At the same time, the IEEE AI Education Large Model Standards Group (P3428), led by Squirrel AI, was established. The first expert working group includes Professor Huang Ronghuai of Beijing Normal University, Professor Xu Bin of Tsinghua University, and Professor Gu Xiaoqing of East China Normal University.

And Li Haoyang, founder and chairman of Squirrel Ai, serves as the chairman of the working group, which is enough to show his representativeness in the industry. We know that once industry standards are established and the application of large models in the education field is further standardized, the industry's development level will be raised to a whole new level.

In the past year, we can see that large-scale models have been implemented in the field of education. Now with the deepening of implementation, the progress of Squirrel AI may be worthy of attention.

Squirrel AI intelligent adaptive education large model

According to reports, the entire Squirrel AI intelligent adaptive education model is mainly divided into three layers: the underlying data layer, the model layer, and the application layer, which reflects its three major characteristics.

The first is the data level.

Squirrel AI's years of accumulation in the education scene have formed a rich underlying data layer. It includes nanoscale knowledge maps of all subjects, massive learning materials, video explanations, assessments and question banks, as well as student learning history data.

It is worth mentioning that the student learning history data here is the entire process of 10 billion learning behaviors of 24 million students accumulated by Squirrel AI in the past, such as learning progress, learning effects, learning paths, and even the time spent in a daze during the learning process.

With personal learning data, on the one hand, we can accurately grasp the user portrait of each student, and on the other hand, we can understand the learning situation of students of the same age, which will also help in the planning and arrangement of the overall knowledge points of the subject.

This also determines that the large Squirrel AI model is fundamentally different from other traditional large language models, such as the GPT series. They will be more personalized and tailor the medicine to the student's learning situation.

Then at the model layer , when talking about large models in the past, it was nothing more than technologies and applications such as multi-modality, LLM, and Agent. However, Squirrel Ai combines the knowledge graph (KG) & retrieval enhanced generation (RAG) of the large model, which is worth mentioning. said.

Simply put, the KG and RAG enhancement technology combined with large models can quickly find the relationship between knowledge points, between knowledge points and questions, and between questions and children's abilities.

If children are given practice questions that are more closely related to knowledge points, their learning efficiency will be higher. Different students have different mastery of knowledge points, so for specific knowledge points, questions that are strongly related to that knowledge point need to be pushed.

Li Haoyang gave a concrete description. At present, they can test 1,000 knowledge points with 25 questions, while the highest on the market can only test 100 knowledge points with 25 questions, and they can also break down each problem in detail. Go through the steps to explain step by step.

In addition, Li Haoyang also emphasized that Squirrel AI's LAM (Large Adaptive Model) intelligent adaptive large model applies the MoE (Mixture of Experts) hybrid expert system to obtain better prediction performance by combining multiple models. Reduce the inference cost of the model.

At the application layer , it covers multiple scenarios such as learning recommendation, learning interest inspiration, habit cultivation, emotional intervention, and learning path planning.

For example, when providing tutoring to students, it can cover multiple scenarios such as preview, review, test preparation, homework guidance, etc., and provide students with more accurate responses and incentives.

In addition, we can also give positive feedback and incentives to children based on their reactions during the learning process to improve students' emotions and help solve some of their psychological problems.

It can be said that it has some of the other large model products, and it also has some that do not.

How to evaluate?

To sum up, from the data to the application layer, they all revolve around one thing in common: personalization.

Looking at product functions, Squirrel AI intelligent adaptive education large model has all the functions, and compared with the large language large model, the technical implementation is more complex. In terms of application scenarios, it is also better able to grasp students' personal learning situations to achieve personalized applications.

As for why it can achieve such an effect, Squirrel AI’s three major layouts and choices deserve attention:

Nanoscale knowledge splitting, MCM systems, large model throughput data types.

The first is the splitting of nano-level knowledge points. For example, in a mathematical scenario, focusing on the major knowledge point "finding the unknown term in addition and subtraction" alone, you can split it into "two-step equations for fractions with the same denominators - the unknowns are subtractions" and "two-step equations for fractions with different denominators" "The unknown number is an addend" is a level of knowledge.

This is equivalent to building a huge knowledge network in the entire learning system, which can help students locate their learning situation more accurately and plan their learning path more accurately.

The second is the MCM system. The first M is Model of thinking, which is the thinking model, the second C is Capacity, which is the learning ability, and the third M is Methodology, which is the learning methodology. The MCM system builds a model based on learning thinking, abilities, and methods, so that it can more scientifically evaluate students' learning status and ability levels in real time.

After loading the large model capabilities, it can predict students' ability level and time spent on knowledge points that they have not learned, so as to launch personalized solutions.

The third is that the throughput data types of large models are different. As mentioned earlier, data types are more complex and the capabilities that large models can provide are more diverse.

Having said so much, in fact, whether it is technology integration or data selection, there are underlying reasons behind it.

That is the choice of "intelligent adaptive education" , which is also the most talked about keyword in the entire press conference.

What is intelligent adaptive education? Simply put, it is based on AI, big data, Internet of Things and other technologies, combined with a large amount of user data, to provide suitable education forms based on differences in individual learning processes, so as to achieve personalized education and teach students in accordance with their aptitude.

Currently, traditional giants including Pearson Education, McGraw-Hill, Wiley, and HMH are also making plans.

As one of the earliest companies in China to implement this concept, Squirrel AI has already developed the core intelligent adaptive learning engine architecture.

△ Intelligent adaptive learning engine architecture

According to Zhou Wei, co-founder and CEO of Squirrel AI, Squirrel AI intelligent adaptive education uses a three-layer architecture to build knowledge maps, learning strategy architecture, content maps, teaching processes, and conduct data analysis, recommendations, and algorithms. , and finally form a complete closed loop of teaching.

With the arrival of large models, the realization of personalized education itself is a rare opportunity, so the integration of Tongzhi adaptive technology has become timely.

According to Squirrel AI's understanding, the fully automatic standard of intelligent adaptive technology can be split according to the evaluation system of autonomous driving.

△ Intelligent adaptive education grading chart according to intelligence level L1-L5

Li Haoyang believes that most mainstream players in the industry are still in the L2 development stage . That is, based on the student's learning situation, intelligent adaptation of question data quantity and question difficulty is realized, which is assisted driving. And they have achieved a level of 40% in the L5, fully autonomous driving stage.

When talking about the final form of AI transforming education, their ultimate goal is to realize virtual teachers, which will help generate more video explanations and help cultivate innovative talents that society needs. In this way, while pursuing personalized education, universal education can truly be achieved.

At least now, it's not that far-fetched.

At the press conference, Squirrel AI introduced that by 2023 Squirrel AI has completed the business layout of 2,000 offline intelligent learning machine stores and is one of the largest AI learning machine brands in retail stores in the country.

As large models are further deployed and put online on their learning machine products, they can quickly bring industrial value.

What changes will big models bring to education?

In the past year, large model technology has rapidly iterated, and AI application possibilities have continued to expand. Especially in the field of education, it is considered to be one of the scenarios where large models should be implemented.

Many companies have planned around this scenario, including the artificial intelligence teaching assistant Khanmigo launched by "Khan Academy"; Duolingo Max with built-in GPT-4 launched by Duolingo.

To sum up, there are several main paths for large-scale model education.

One is similar to ChatGPT, which solves the problems of students or teachers on the web/APP side . Large models act as learning assistants through natural language interaction. This is most common in language learning scenarios, such as Duolingo, and the recent high-profile startup Speak.

The other is to deploy large models on the original learning hardware . This kind of natural interaction scenario also has a user and data basis. The application iteration speed may be faster, but the technical requirements are higher. The existing forms include learning machines, educational robots, dictionary pens, etc.

Under the current development situation where a hundred flowers are blooming, it is not difficult to see that hardware forms and software functions tend to be the same and gradually become homogeneous.

The essence of enterprise innovation still lies in the empowerment of AI. This is also the core for enterprises to maintain their competitiveness in the new wave of technology. Once the large model is deployed, more innovative applications can be implemented, user data brings faster iterations, and marginal costs will gradually decrease.

However, on the user side, in addition to explicit functions such as dialogue, more core capabilities of large models are embedded in the product, such as formulating learning plans based on user habits and learning progress, etc., but users often do not perceive it in a short period of time.

This is also a major challenge in the current education scenario.

This requires companies to educate consumers on the one hand, and on the other hand, insist on long-term and continuous technological investment and build corporate barriers to withstand the treacherous external environment. When new technologies come, they can be quickly utilized in conjunction with existing scenarios to serve users.

Squirrel AI, which has deployed technology for nearly ten years, is an example.

But in the end, no matter how large models are applied in education, the general direction has been determined.

That is personalized education and teaching students in accordance with their aptitude.

In fact, this is a goal that the education industry has widely accepted and has always wanted to achieve. As early as the 1980s, the famous educational psychologist Benjamin Bloom proposed the "two sigma" theory in his educational experiments, which proved its correctness. ——

The average performance of the teaching group that received one-on-one tutoring was significantly better than the traditional 1:30 teaching method.

Nowadays, the personalized technical characteristics of large models themselves have brought new possibilities to the reform of the education industry.

There is no doubt that with the deepening of large-scale model layout and the establishment of more subdivision standards, the development of the education industry has reached a key historical node.

Click here ???? Follow me and remember to star~