AI’s Helping Students Improve Their Scores In Less Than 24 Hours
In the last few years, AI-enabled online learning platforms are increasingly popular around the world for parents who want their children to gain an edge on standardized tests or attain mastery of a particular subject matter. In Asia, multiple platforms compete for children’s after school attention. With over 1 million subscribers in South Korea in a short period, Santa for TOEIC, an AI-enabled mobile and web-based learning platform, quickly became popular with its highly educated generation Z.
South Korea is one of the most educated countries in Asia. As English becomes ubiquitous in business around Asia, more children feel the need to score well on the standardized English language test, TOEIC, to earn entry to universities or apply for better job opportunities. When Santa for TOEIC released its initial application, users quickly realized how it helps them stay on track with their learning. With just 20 hours, users can increase by 165 points out of 990 points within the first 20 hours of studying.
I had the pleasure of speaking with Young-Jun (YJ) Jang, CEO of Riiid. YJ started Riiid to partner with elite research institutions in Seoul, South Korea, to apply cutting-edge AI algorithms to solve online learning problems.
In 2014, YJ started Riiid as a research and development effort to explore opportunities to use AI in various industries. Quickly, he realized that there’s an urgent “One Size Fits All” problem or the adaptive learning problem in education, particularly in Asia. Most of the algorithms in the space tried to solve this problem were all rule-based algorithms. These rule-based algorithms were tailored at the individual level, but they didn’t truly adapt to the learner and adjust according to how they learned.
YJ says, “We also had a problem in terms of scalability and advocacy. When you hire a bunch of domain experts to create and analyze each content to be personally distributed to the users in an adaptive learning system, it’s a lot of money spent on the companies and startups on personalization. The fact that a person is involved, there are also ethical issues as well as privacy issues. So, I spoke at length about this problem with computer scientists at UC Berkeley and Stanford. Eventually, one of them pointed me to machine learning. I quickly realized that these problems could be best solved by machine learning or deep learning algorithms.”
As YJ worked with his team of technologists and researchers, he quickly realized that AI is a fast-evolving field. When researchers developed better algorithms, his company had to take bold steps to test, implement, and improve the existing algorithm. Not only that, as the company collects more data, but there are also opportunities to define smaller problems within the bigger scope of the problem to understand the nuances. The iteration of fine-tuning the algorithms as smaller problems were solved to inform the larger problem’s solution is what allows the predictions to continue to improve.
YJ says, “The first algorithm that I was interested in was the collaborative filtering algorithm. Over time, we changed our algorithm, and we got to a point where we can predict whether the user gets the next question right or wrong. We also can predict which answer choice they can pick with 90% accuracy. We did a lot of experimentation, inspired by different learning algorithms, such as the BERT algorithm from Google. These days, most deep learning models rely on a specific format of input data. But, we are progressing into a more meta-learning environment where the reinforcement learning model can define its own problem instead of focusing on finding the solution to a pre-defined problem.”
One example that YJ recently encountered at Santa for TOEIC was that within one session, sometimes, there are different users logged in during the same day.
YJ says, “We weren’t aware of this. We thought the user was the same, so that we can use the data we collected previously.
