The promise of digital technologies transforming learning has stayed a mirage for decades. However, it took many more decades to finally see it in practice, thanks to widespread smartphone adoption and the rise of AI/ML-powered technologies like adaptive learning. 

Duolingo is a very familiar name when it comes to language learning. The language learning app, started by Carnegie Mellon University professor Luis von Ahn and his post-graduate student Severin Hacker in 2009, has received numerous awards and honours and has over 500 million users on the platform learning over 100 courses across 40+ distinct languages. 

To understand how Duolingo creates a high-quality learning experience by combining human expertise with AI in a way that leverages the strengths of each, we reached out to Klinton Bicknell, Head of AI at Duolingo. Klinton, who leads the AI Research at Duolingo, has deep expertise in Linguistics, Cognitive Science and Interdepartmental Neuroscience. Klinton is PhD holder in Linguistics and Cognitive Science from the University of California, San Diego and winner of numerous honours and fellowships.

Klinton, What do you think is the role that AI and ML will play in the learning space in the coming year?

For companies like Duolingo, with a large amount of data about how users learn, AI will play a significant role in figuring out the best way to teach for everyone and personalizing the teaching to each learner. AI will also be important in helping keep each learner motivated to keep coming back. 

Separately from teaching, but still in the education space, AI will also play an increasingly prominent role in assessment and testing, as it does in the Duolingo English Test, where AI is used to develop, administer, and score each test.

Can you explain how AI and ML technology is being used in the Duolingo app?

We use AI and ML throughout our main language learning app. To name a few uses of AI :

• Personalized, adaptive lessons: By using AI, we are able to customize lessons to each learner, particularly by understanding at which points they need more practice and then generating sessions accordingly. Once learners have begun a course, AI works in the app's background to model the words, concepts, and skills they have mastered and what they still need to practice.
• Adaptive placement tests on Duolingo: Before starting a Duolingo course, a user is asked to select whether they're a complete beginner or have had prior experience with the language they want to learn. If they choose the latter, they are directed to take a placement test to determine their proficiency before beginning the course. The test adapts based on how a user does as they advance through the test; if a user is consistently correct, the placement test will ask more challenging questions, and vice versa if a user starts to make mistakes.
• Smart Tips: Smart tips are explanations that appear in a lesson after a learner makes certain types of mistakes, like putting words in the incorrect order in a sentence. Because mistakes are a natural part of language learning, we designed smart tips to give personalized feedback to help you learn from your mistakes.

Can you briefly explain the AI systems and tools being used?

We build many in-house AI systems for many different tasks at a high level. Some of these are straightforward linear models implemented directly in python or java, and others are sophisticated deep learning models built with PyTorch.

For example, the primary model we use to personalize lessons is one we affectionately call Birdbrain. Birdbrain is a machine learning model that predicts how likely a user is to get right each possible exercise we might serve them, based on all the data we have about exercises that the user has answered in the past.

Formally, the current version of Birdbrain is implemented with artificial neural networks, specifically a version of a recurrent network. Duolingo learners answer around a billion exercises daily, so Birdbrain is trained on a large amount of data. It knows that learners who get a particular exercise wrong are also likely to get some other exercise wrong, etc. We then use this model to personalize what exercises we serve a learner in several ways. 

Additionally, we use the predicted probabilities of a learner getting each exercise correct to serve learners exercises that aren't too easy for them – which would be boring and also not very useful for learning – or exercises that would be too hard for them – which may be discouraging and also not very useful for learning. By keeping exercises in the intermediate difficulty range, we can maximize learning efficiency and keep our learners engaged.

How do you see the future of learning changing with the advancement of technologies?

It's a very interesting question! By combining AI with the massive amounts of data produced about learning by modern educational software like Duolingo, we should be able to substantially optimize and personalize learning trajectories to be more efficient and more engaging – like we are doing at Duolingo with Birdbrain. As technology advances, the rate at which people can learn new topics or skills may increase substantially. Having these high rates of learning available in free apps like Duolingo may give learners around the world access to high-quality, free, and effective education.