Where Machine Learning Goes Wrong (And Right) for Education

Vibhu Mittal, CEO, Edmodo

The rise of inappropriate children’s videos on YouTube has put Machine Learning at the center of the conversation for Ed Tech. Machine Learning is clearly a powerful tool that can enrich lives and save time, but YouTube’s example has shown how educational services need to be careful when using it.

Machine Learning is one of the reasons we love services and apps like Facebook, Spotify, and Netflix. They create personal mixes and recommendations just for us, saving time we would have spent searching for the right content. But, it’s not perfect. There are still Facebook articles that make you cringe, Spotify playlists with songs you always skip, and Netflix recommendations that miss the mark entirely.

MDR Machine LearningThe problem with Machine Learning is not the algorithm itself but what the algorithm is optimized for and what its goals are.

In the case of the knock-off children’s videos on YouTube, these creators are undoubtedly optimizing their content for revenue or, in YouTube terms, views.

These videos, and the algorithm itself, are optimized for attention, not learning.

All it takes is a decent amount of traction in viewership for YouTube’s algorithm to pick up the video and suggest it to other, similar users. These video creators are using every trick in the book to get the algorithm on their side too, filling their video titles with keywords that favor recommendation. And without a human hand to guide or curate these results, things quickly go off the rails.

Recognizing this, YouTube announced on December 5th they will be hiring thousands of human beings to review videos.

What’s the solution?

At Edmodo, we believe Machine Learning has the potential to increase learning outcomes—when it’s applied to the right problems with the right controls. As we’ve begun developing machine learning technologies, we’ve learned a lot about what will make it successful for education.

This year, we began an experiment called AskMo. We took millions of videos that teachers shared with their classes, anonymized them, and built a recommendation engine around them. This circumvents a lot of the normal issues with machine learning. Since these videos were shared by teachers, they’re already more credible than just “popular” videos, and they’re further vetted by teachers sharing them with their classes.

Even then, we found that videos some teachers deemed inappropriate quickly rose in the ranks. These videos weren’t inherently disturbing but spoke on topics that many didn’t feel comfortable sharing with students: cyberbullying, dangerous animals, and artificial body parts.

Even when being careful about video content, there are still plenty of questions that need answering:

  • How do we stop inappropriate videos from dominating in recommendations?
  • Who gets to define what is or isn’t appropriate, when the definition varies by age, culture, and location?
  • How do we optimize for learning instead of attention?
  • How do we quantify learning, and what metric should we use?

These are important questions to ask and ones we believe the Ed Tech community will need to answer together. It’s going to take a lot of iteration and refinement before we get there.

But there is good news. Machine learning has plenty of uses to impact learning. And we’ve discovered one of the strongest: finding a student’s passion.

Teachers know that if you can spark passion in students, they become lifelong learners in that subject. Usually this is from a good match of teacher and student—or serendipity with a specific lesson.

We want to use data we have, including over 600 million shared resources, to help each student. If we can give students the chance to learn about a wide variety of subjects, we can help them find and grow their passion faster and more precisely than ever before, opening students to areas they might not have discovered otherwise. With AskMo, for example, we can help students explore areas they don’t yet know or fully understand. They can start with an inkling of interest in any subject and click through related topics. It functions like a funnel—the students’ own interest gently pushing them toward the topic they’re most passionate about.

Before the advent of the Internet, letting students explore their own interests usually meant getting lessons through school or private tutors, which meant a lot of time and money invested into something that the student often resisted. This is certainly disappointing for teachers and parents, and it can make some students feel aimless and lost.

With Machine Learning, students can breeze through different topics, gravitating toward the one that sparks that passion. Since there’s no cost or major time investment, this saves both parents and teachers valuable time and money, which they can use to help encourage a verified interest.

Machine Learning has a long way to go, but it can unearth a passion that transforms a student into a lifelong learner. And, at Edmodo, we’re eager to support the Ed Tech community in finding the balance for Machine Learning—the right ways to use it, with the right objectives, and how to mitigate its potential problems.

Vibhu Mittal is CEO of Edmodo, a global education network that connects teachers, students, parents and admins, and provides access to millions of of shared resources to improve learning. Vibhu has a deep commitment to education and has helped create various educational and learning systems, including Google Scholar, Google Translate, and Root-1, a mobile platform for teachers to share interactive, educational content with students, which was acquired by Edmodo in 2013.

Have something to add to this article?