Jennifer Turner’s algebra classes were once sleepy affairs and a lot of her students struggled to stay awake. Today, they are active and engaged, thanks to new technologies, including an artificial intelligence-powered program that is helping her teach.
She uses the platform Bakpax that can read students’ handwriting and auto-grade schoolwork, and she assigns lectures for students to watch online while they are at home. Using the platform has provided Turner, 41, who teaches at the Gloucester County Christian School in Sewell, New Jersey, more flexibility in how she teaches, reserving class time for interactive exercises.
“The grades for homework have been much better this year because of Bakpax,” Turner said. “Students are excited to be in my room, they’re telling me they love math, and those are things that I don’t normally hear.”
For years, people have tried to re-engineer learning with artificial intelligence, but it was not until the machine-learning revolution of the past seven years that real progress has been made. Slowly, algorithms are making their way into classrooms, taking over repetitive tasks like grading, optimizing coursework to fit individual student needs and revolutionizing the preparation for College Board exams like the SAT. A plethora of online courses and tutorials also have freed teachers from lecturing and allowed them to spend class time working on problem solving with students instead.
While that trend is helping people like Turner teach, it has just begun. Researchers are using AI to understand how the brain learns and are applying it to systems that they hope will make it easier and more enjoyable for students to study. Machine-learning powered systems not only track students’ progress, spot weaknesses and deliver content according to their needs, but will soon incorporate humanlike interfaces that students will be able to converse with as they would a teacher.
“Education, I think, is going to be the killer app for deep learning,” said Terrence Sejnowski, who runs the Computational Neurobiology Laboratory at the Salk Institute for Biological Studies in La Jolla, California, and also is the president of the Neural Information Processing Systems Foundation, which each year puts on the largest machine-learning conference in the world.
It is well established that the best education is delivered one-to-one by an experienced educator. But that is expensive and labor intensive, and cannot be applied at the scale required to educate large populations. AI helps solve that.
The first computer tutoring systems appeared in the 1960s, presenting material in short segments, asking students questions as they moved through the material and providing immediate feedback on answers. Because they were expensive and computers far from ubiquitous, they were largely confined to research institutes.
By the 1970s and 1980s systems began using rule-based artificial intelligence and cognitive theory. These systems led students through each step of a problem, giving hints from expert knowledge bases. But rule-based systems failed because they were not scalable — itwas expensive and tedious to program extensive domain expertise.
Since then, most computer teaching systems have been based on decision trees, leading students through a preprogrammed learning path determined by their performance — if they get a question right, they are sent in one direction, and if they get the question wrong, they are sent in another. The system may look like it is adapting to the student, but it is actually just leading the student along a preset path.
But the machine-learning revolution is changing that. Today, learning algorithms uncover patterns in large pools of data about how students have performed on material in the past and optimize teaching strategies accordingly. They adapt to the student’s performance as the student interacts with the system. Bakpax asks teachers to notify parents how their children’s data will be used, and parents can opt out. But Bakpax and other companies say they mask identities and encrypt the data they do collect.
At its core, Bakpax is a computer vision system that converts handwriting to text and interprets what the student meant to say. The system’s auto-grader teaches itself how to score.
“Instead of handing your homework in, you just take a picture of it on your phone, and a few seconds later we can tell you what you got right and what you got wrong,” Ferreira said. “We can even tell you what the right answer is for the ones you got wrong.”
The system also gathers data over time that allows teachers to see where a class is having trouble or compare one class’ performance with another. “There’s a lot of power in all this information that, right now, literally is just thrown in the trash every day,” Ferreira said.
Not surprisingly, machine-learning solutions are making their way into the test preparation market, a multibillion-dollar global industry. Riiid, a Korean startup, is using reinforcement learning algorithms — which learn on their own to reach a specified goal — to maximize the probability of a student achieving a target score in a given time constraint.
Riiid claims students can increase their scores by 20 percent or more with just 20 hours of study. It has already incorporated machine-learning algorithms into its program to prepare students for English-language proficiency tests and has introduced test prep programs for the SAT. It expects to enter the United States in 2020.
Still more transformational applications are being developed that could revolutionize education altogether. Acuitus, a Silicon Valley startup, has drawn on lessons learned over the past 50 years in education — cognitive psychology, social psychology, computer science, linguistics and artificial intelligence — to create a digital tutor that it claims can train experts in months rather than years.
John Newkirk, the company’s co-founder and chief executive, said Acuitus focused on teaching concepts and understanding.
Newkirk likens AI-powered education today to the Wright brothers’ early exhibition flights — proof that it can be done, but far from what it will be a decade or two from now.