The AGE of AI: In the age of AI, AI-driven adaptive learning has emerged as a novel paradigm, transforming the landscape of learning. At the recently concluded KDD 2024, Squirrel Ai took the stage to share with everyone what true personalized learning should look like.
SHANGHAI, Oct. 10, 2024 /PRNewswire/ -- As the new school year kicks off, the new buzzword has gone viral -- "painless learning."
Nowadays, AI has rapidly integrated into the study and lives of university students at an unprecedented pace, and more and more people are beginning to experience what it feels like to shift from "painful learning" to "painless learning." But is AI only changing the lives of university students? Absolutely not. Recently, a news story broke about AI offering creative ideas to a young boy in Chengdu, China, to mass-produce hand-drawn posters, frantically catching up on his summer homework with the start of the school year approaching.
Whether you embrace or hesitate towards it, the undeniable truth is that AI is profoundly impacting the field of education. From elementary students to university students, none can escape this overwhelming wave. What is it about AI that makes it blend so seamlessly with education? The reason lies in the fact that education should be tailored to each individual, but teachers cannot always give attention to every student. AI, however, completely breaks through this limitation in reality. Moreover, thanks to cutting-edge technologies like LLM (Large Language Models) and AIGC (AI-generated Content), traditional adaptive learning has gradually evolved into more intelligent "adaptive learning." Some representative educational institutions have seized the opportunity and are investing in adaptive learning. For example, Duolingo Max employs role-playing techniques, allowing students to converse with AI through simulated scenarios.
Khanmigo uses a platform with personalized LLM dialogue, asking continuous questions to help students build knowledge while providing personalized teaching, ultimately resolving their doubts.
Andrew Ng's Coursera, based on open online courses, offers online adaptive learning courses and a blended learning model. Clearly, adaptive learning is becoming a consensus in the AI education sector. Coincidentally, at the recently concluded ACM KDD 2024, the closing ceremony's major roundtable discussion focused on GenAI + Education.
During the roundtable, participants including Professor Nitesh Chawla (AAAS/AAAI/ACM/IEEE Fellow) from the University of Notre Dame, Professor George Karypis (IEEE Fellow) from the University of Minnesota, Dr. Joleen Liang, Co-founder of Squirrel Ai, and Professor Ricardo Baeza-Yates (ACM/IEEE Fellow) from Northeastern University, dissected the outlook for the future of GenAI and examined potential innovative use cases.
Additionally, Squirrel Ai's R&D team presented a paper at the main session of KDD, exploring the application of large language models (LLMs), specifically educational models, in time series analysis. They also hosted a workshop and delivered some of the keynotes during GenAI Day.
As the longest-running and largest international conference in the field of data mining, and an A-class academic event in China, KDD annually attracts thousands of leading scholars and business leaders from around the globe. It is a rare occasion for businesses to participate in academic discussions at such an international top-level gathering. Squirrel Ai, as of the few companies invited to participate and present a paper, demonstrated its proven expertise in AI-driven learning. So, what do these industry leaders think about the current development of AI?
What True AI Personalized Learning Looks Like
Years ago, when AI was still nowhere as ubiquitous as it is today, Squirrel Ai faced numerous questions – why incorporate AI into education? What can it do? The answer is we use AI to provide students with a true personalized learning platform and content. Today, everyone is considering how to maximize the value of GenAI. If the goal is just to make a quick buck by establishing a company in a short time, this mindset is actually very dangerous. Instead, leveraging this technology wisely, for example, in data analysis to enhance student learning, rather than merely churning out content – presents a significant opportunity. Currently, we are shifting the focus to human-machine interactions.
As we embark on interactions fueled by language and multimodal capabilities, we find ourselves at the dawn of a robots-led industrial revolution. However, we are still at a very early stage of using AI in education. Despite AI, MOOC (Massive Open Online Courses), online and remote learning, as well as speech and semantic recognition, the essence remains rooted in conventional learning methods. They fail to deliver truly personalized learning experiences. What is personalized learning? Many assume that it is about typing a question into ChatGPT and getting back an answer, but this is far from the truth. Dr. Joleen Liang indicated that most users, even companies, are yet to understand the nature of personalized learning. Hence, we have a long journey ahead to demonstrate what a truly personalized learning and AI-driven adaptive learning technology entail. As a minimum requirement, it should be one that allows users to understand, experience, make mistakes and re-experience. The next step involves the integration of AI, personalized learning, AI-driven adaptive learning systems, multimodal capabilities, and large models.
AI-driven adaptive learning not only helps students achieve higher scores, but also enables them build a broader range of skills within the same learning timeframe. Envision a scenario where students of varying ages sit around a table, without a teacher. They collaborate on group assignments or work together to address challenges. Students learn autonomously with interconnected devices, while teachers take the role as assistants, data analysts, emotional supporters, and mentors. This vision of AI-driven adaptive learning models represents the future of education.
GenAI makes a difference in education
The shortcomings of K-12 education are widely acknowledged. As we enter a technological renaissance, schools and parents aspire for more intelligent and personalized educational experiences for their children. Can GenAI provide innovative applications in the K-12 education sector? If so, the future of education will be revolutionized, offering more value to students and educators alike. The characteristics of GenAI address the needs of K-12 education.
Personalized learning: the common metric for LLMs in education
The question becomes, is there a common standard for evaluating the application of GenAI (Generative Artificial Intelligence) in education, and what should that standard be? In 2022, a report released by Google mentioned three major trends in future education, one of which is "personalized learning" with adaptive capabilities.
Personalized learning underscores student autonomy, where learners take charge of their educational journey unhindered by external influences. The approach is paramount to fostering important skills and competencies. Throughout the process, learners work with peers and seek guidance from educators. In a 2021 report, EY likens the progression of education to the stages of autonomous driving based on the use of intelligent technologies, categorizing it into six levels (L0 to L5). L0 represents traditional human-led instruction without automation. L1 reflects early digital integrations, including live and recorded lectures. L2 encompasses supportive tools such as photo-based question-solving approach, while L3 to L5 mark the full integration of AI throughout the learning journey for personalized experiences. At the pinnacle of this evolution stands L5 – "fully intelligent adaptive education," where AI assumes the primary role in instruction, eliminating the need for any external resources. Squirrel Ai provides a model example.
Achieving L5: Fully AI-based adaptive learning
In the K-12 education sector, Squirrel Ai, a company with over ten years of expertise, understands that GenAI can dynamically adjust learning content and difficulty by analyzing students' learning data. This approach offers a learning path tailored for each individual. The data-driven teaching method enhances learning efficiency and effectively fills in students' knowledge gaps, ensuring their holistic development. Here's how powerful GenAI strengthens students' learning outcomes:
- Knowledge Point Breakdown: To build learning skills, it is necessary to understand and correlate the knowledge points within the system. Squirrel Ai achieves this goal with Socratic-style guided questions, which prompts students to think actively and deepen their comprehension, creating a more efficient and personalized learning environment. Squirrel Ai uses its unique database containing 10 billion learning behavior data points collected from the entire learning journeys of 24 million students. The data is fed into the LLM for iterations in recommendation algorithms and deep knowledge tracing. Consequently, the model adeptly captures complex relationships and patterns within the data, quickly identifying connections between knowledge points, questions, and students' skills. By doing so, the model can generate accurate learning profiles and offer tailored, personalized services.
- Error Analysis: Squirrel Ai's new Large Adaptive Model introduces an intelligent analysis function for draft paper content. The feature allows deep analysis of every step in a student's problem-solving process, pinpointing errors in question comprehension, logical reasoning, calculation, and handwriting transcription. This ensures both students and teachers can identify issues quickly and accurately, and address weaknesses accordingly. This is made possible by the significant leaps in regular assessment efficiency and accuracy achieved by the new version of the multimodal large model, especially in subjective question grading. For example, it can accurately assigns scores and provides detailed explanations of deducted points, offering a comprehensive insight into students' learning outcomes.
- Human-Machine Interaction: In terms of intelligent human-machine interaction, the new multi-modal adaptive engine has reached a new height. It supports both text and voice interaction modes, covering over 100 interactive dialogue scenarios. More thoughtfully, it possesses high-precision emotion recognition capabilities. With the feature, if a student experiences emotions like confusion, happiness, or distraction, the model can accurately detect them and promptly provide personalized feedback and encouragement in real-time. Furthermore, to create a more comprehensive and immersive audiovisual learning experience, Squirrel Ai has upgraded its intelligent scanning pen and headphones in a move to build a complete AI-based adaptive learning ecosystem. Its flagship product —the Smart Instructor S211 Egret—has revolutionized the traditional setup by using the unique digital paper technology, offering original color, color ink, and water ink display modes, paired with a high-definition camera to capture learning details in real-time. It sets a new industry benchmark in user experience and health considerations.
Squirrel Ai's Large Adaptive Model LAM
So, how does the multi-modal adaptive engine LAM proposed by Squirrel Ai work behind the scenes?
The model's architecture consists of three key components: a knowledge graph, a recommendation engine, and Retrieval-Augmented Generation (RAG). The recommendation engine contains different planners and agents, including both short-term and long-term path planning. Other components include sentiment analysis, path tracking (as each student has their own learning path), feedback, and summarization from the large model. Additionally, the AI system also incorporates two essential elements: parent goals and student profiles.
Within the intelligent agent architecture, the entire agent is called the "adaptive agent." Its architecture houses an array of agents: data analysis agent, teaching autonomous driving agent, instructional Q&A agent, reasoning agent, and reading comprehension agent. Discipline-specific agents for physics, English, and science, work together to engage with students and provide them with personalized content recommendations. To facilitate the efficient coordination of these multi-faceted agents, the research team has developed a three-tiered adaptive engine. Tier 1 encompasses Goal, Learning Map, Content Map, and Root Cause Analysis. These dynamic elements set not just static targets, but also adapt to students' pace, progress, and data insights in order to ensure recommendations are always optimized and relevant. The Root Cause Analysis, in particular, provides pinpoint accuracy in diagnosing and addressing student misconceptions.
Tier 2 focuses on learning records. The AI system keeps track of and evaluates all students' learning materials. Tier 3 is realized by seamless interaction between the students and the AI system. Real-time data from these engagements is fed into the AI-based adaptive engine, where the AI system performs computation and analysis. Based on these insights, the AI system dynamically recommends tailored learning content, including knowledge, MCM (methodology, capability, mindset) skills, and practical applications. In addition, Squirrel Ai's engine incorporates other key technologies, including the world's first "Nano Level Granularity Knowledge Graph". Recognizing that learning objectives vary widely across grades and subjects globally, Squirrel Ai's research team has sorted these objectives into refined layers for developing algorithms. To cite an example, an objective about "addition and subtraction of fractions" can be broken down into second-level learning objectives (e.g. addition of fractions, subtraction of fractions, simple fraction calculations, multi-step fraction calculations). Third-level learning objectives with finer granularity can be derived from the second level, as shown in the figure below. Ultimately, one objective can be divided into nine layers.
Squirrel Ai's Large Adaptive Model (LAM) also comes equipped with a prediction engine capable of estimating up to 100 learning objectives to be met based on a student's 10 hours of study. The estimates dynamically change as the student progresses. Additionally, the recommendation engine within Tier 1 dynamically adjusts learning objectives based on real-time data. As shown in the figure below, the engine works as a root cause tracing system. Taking Grade 10 as the highest and Grade 7 as the lowest, from bottom to top, the green line refers to the learning objectives, knowledge, and skills that students have mastered. Suppose a student is in Grade 10 and faces difficulties in understanding a concept. The AI system will then trace the source of the question and suggest that the student review specific courses at the Grade 7 level. Only after ensuring that he has mastered the knowledge points, the AI system will resume his learning path.
Consider a scenario where three students each demonstrate an 80% mastery of learning objectives. Despite this shared overall proficiency, Squirrel Ai's system reveals distinct strengths and weaknesses among them, as indicated by the distribution of their remaining 20% of knowledge. This underscores the importance of a robust problem tracing system, which empowers educators and students alike to address individual learning gaps.
The inclusion of the MCM system demonstrates Squirrel Ai's commitment to equipping students with the skills demanded by various industries and professions. From language to physics to mathematics, each discipline requires a unique mix of skills, and Squirrel Ai's MCM training ensures that students are well-prepared to excel in their chosen fields.
Squirrel Ai's Large Adaptive Model LAM stands apart from conventional LLMs. Taking ChatGPT for example, the system provides answers to queries and maintains the query history. However, ChatGPT does not recommend personalized learning content and cannot understand an individual's learning history and proficiency. By contrast, Squirrel Ai's LAM leverages learners' historical data, unique learning behaviors, and comprehensive assessments to offer tailored recommendations and detailed performance reports, paving the way for more effective and efficient learning journeys.
Sustained model iterations
Within the multi-modal adaptive engine, Squirrel Ai's R&D team introduced a new algorithm at this KDD conference—time series analysis. In their paper titled "Foundation Models for Time Series Analysis: A Tutorial and Survey," the team systematically discussed the application of large models in time series analysis.
Specifically, the paper examines the application of models in education, including their role in educational time-series data analysis. Time series analysis is increasingly being applied in educational scenarios, allowing for the prediction of student learning progress and the optimization of teaching strategies through the analysis of student behavior data, test scores, and learning habits. By pre-training on large-scale educational datasets, GenAI for time series analysis can better capture complex temporal dependencies and nonlinear relations, enabling optimized prediction and classification performance across multiple educational contexts. The paper also identifies several potential avenues for future research, such as leveraging multimodal data (text, images, speech) to further enhance model generalization capabilities and using self-supervised learning techniques to reduce reliance on labeled data. These endeavors not only help improve the accuracy and efficiency of time series analysis, but also provide a solid technological foundation for personalized and intelligent adaptive education.
Squirrel Ai's engagement with top scholars from around the world at the KDD conference reflects the company's comprehensive strengths spanning academia, research, and business. As the age of AI fully unfolds, we hope to see more Chinese companies participating in such top-tier summits.
AI is approaching the ideal educational scenario
The famous educator Vasily Sukhomlinsky once said, "Among the thousands of students who have passed through my hands, it is not the model students who left the deepest impression, but the ones who were unique and different."
In the evolving landscape of AI-driven education, one of the most promising developments is the realization of the long-held dream of "a tutor for every child." This ideal, once distant, is now within reach. At this stage, perhaps AI's greatest contribution to humanity lies not in the delivery of fragmented facts, but in its ability to unlock deeper learning. By the time students graduate, they will leave not just with knowledge and skills, but with a lifelong hunger for learning.
