Adaptive Boosting for Human Learning

I am about to complete the data science immersive program at General Assembly (This week! Very exciting!), but just prior to starting this course, I was a high school math teacher. I’ve taught general education students, special education, and students identified as gifted.

I don’t think it will shock many reading this when I say that a lot of students struggle with high school math. I’ve worked hard and struggled alongside them to try to make it a bit better for them. This year with the pandemic my own kids are learning at home, and I’ve gained new perspective on just how much anxiety it can induce.

The past 12 years of my career have been spent facilitating learning in humans. Now I find myself building skills to facilitate learning in machines, and the parallels are interesting. Midway through this course, I learned about AdaBoost, and I feel compelled to ask:

Why are we so much more patient with machines than with our own human children?

Or, to put it another way,

Can we use the methods of Adaptive Boosting to help kids learn?

First, if you aren’t familiar with AdaBoost (or Adaptive Boosting), it’s essentially the idea that you train an iterative sequence of models, where each model is forced to contend with a larger proportion of the data it got wrong in the previous iteration. There’s a great explanation here by Maël Fabien (graphic also from the same source) :

weighted errors in AdaBoost classification

This fits in pretty well with the “growth mindset” in education:

Over 30 years ago, Carol Dweck and her colleagues became interested in students’ attitudes about failure. They noticed that some students rebounded while other students seemed devastated by even the smallest setbacks. After studying the behavior of thousands of children, Dr. Dweck coined the terms fixed mindset and growth mindset to describe the underlying beliefs people have about learning and intelligence. When students believe they can get smarter, they understand that effort makes them stronger. Therefore they put in extra time and effort, and that leads to higher achievement.

The problem is, this growth mindset is implemented only sparingly in many classrooms, and so many kids simply don’t believe in their capacity to grow in skills and understanding.

Here’s how a typical unit plays out in a high school algebra class. (I’ve chosen solving quadratics as an example, and this is a simplified flow.)

How quadratic equations are usually taught

In this model, most of the grade is often determined by the test at the end. That test covers material the students have forgotten by test day, and they are given (maybe) one retest opportunity to improve their score. Surely there’s a more effective and less soul-crushing approach to human learning!

Boosting for humans, a proposed model (using the same example algebra unit)

How we COULD do it

The key to boosting is iteration, and focus on the areas of previous weakness. I suggest students be allowed to iteratively improve their concept understanding and skills in much the same way. The technology is definitely there — many students have been taking the nwea’s MAP test a couple times a year for some time now. This is an adaptive test that responds to a student’s success or failure on each question to assess their true individual skill level.

What I am proposing is that we integrate this same thinking into the day-to-day life of the classroom. Replace one-size-fits-none repetitive practice sheets with interactive practice that gives each student more of what they specifically need to practice. Replace the punitive end of unit test with yet another chance to go back and grow in areas of continued struggle.

None of this replaces the teacher, of course. Students will need support as they work through new and challenging material. But changing the success metrics, and then actually using those metrics to foster student engagement and achievement, could have a huge impact.

Data Scientist and Mathematician