Stanford Teacher Education Program
Chris Piech, Stanford assistant professor of computer science, turned two teachers into thousands.
Last year, as Piech and his co-teacher were about to start Stanford’s popular Introduction to Computer Science course, pandemic-related lockdowns gave them the opportunity to reach many more learners than just one classroom’s worth.
“I work on how to get education into more people’s hands,” Piech said. “It’s a huge, sticky social problem.” The pandemic enabled him to test a new solution: having thousands of freshly recruited teachers — largely recent graduates — deliver the computer-science curriculum.
The resulting small-group learning sessions became part of Code in Place 2020, which yielded strong learning and retention for thousands of students worldwide.
As part of HAI’s Spring Conference, Piech joined other education experts and scholars in health care and the arts to explain AI’s ability to augment — not replace — critical human work. During the education panel, speakers discussed advances of AI in empowering teachers, leveraging learning-research insights, delivering education in low-infrastructure regions, and scaling workplace learning. (Watch the full conference here.)
AI To Super-Power Teachers
“The number of people in the world who want to teach is proportional to those who want to learn,” Piech said, which motivated his “massive experiment in education.”
AI, in this context, can be used not to replace teachers but to assist and augment. Specifically, Piech’s team has worked on using AI to provide students automated grading and support scalable, more equitable teacher training.
Automating grading is highly challenging. The team found using deep learning to train AI to grade students’ simple coding was ineffective. Having teachers train an AI system to grade yielded marginal improvement. Now, they’re making progress by having AI systems apply learning from past exams to grade new ones. They’ll test the system more fully this year.
On the training front, Piech is overseeing efforts to offer AI-based teacher puzzles and feedback to improve teaching skills. He encourages anyone interested in learning more — or volunteering to teach for Code in Place 2021 — to follow his Twitter.
Avoiding a Legacy of Bad Instruction
According to Daniel Schwartz, Stanford dean and education technology professor, AI-driven educational tools must understand and make use of better learning science.
“New tech often starts by imitating old tech,” he said. “I'm concerned AI may make us more efficient at what is basically not very effective instruction.”
Schwartz instead proposed a more human-centered approach to education that requires understanding exactly how people learn. The last 30 years has seen strong science on better forms of learning, he explained.
For example, observing contrasting cases is critical for new learning but missing from most instruction today. People who are learning what polygons are, for instance, perform much better when shown both positive and negative examples, rather than positive examples only.
Similarly, a Schwartz study examined participants’ ability to apply new knowledge related to memory function after undergoing traditional classroom-type instruction (lecture), gaining experience analyzing new information, or a combination of these. Those who experienced the lecture and analysis work did best. “Analysis led to precise knowledge of key phenomena and contrasting cases, and the lecture generalized the why,” Schwartz said, making that condition superior to lecture or analysis alone.
Also, he noted a study in which participants heard a lecture on ratio and density, then either practiced using the concepts (with the example of organizing clowns in boxes) or were asked to invent a new clown-organizing solution. Those in the invent-a-solution condition did better on learning and application measures than those in the conventional tell-and-practice condition.
“AI and education today just repeat the same instructional information to struggling students, only slower and louder instead of helping them learn it the right way,” Schwartz said. His ideas, including the ones outlined in his book The ABCs of How We Learn, can help those working at the intersection of AI and education use alternative approaches to improve learning.
Pushing the Boundaries of Educational Tech
Globally, we’re seeing a reduction in the number of children who don’t attend school. But COVID-19 sent 1.4 billion students home and, even before the pandemic, low-income countries were struggling to get students the basic primary level skills, says Amy Ogan, Carnegie Mellon professor of learning science.
“The goal is to meet each context where it’s at,” she said. “We’ve seen a global reduction in the number of children out of school but face a learning crisis because kids across countries still struggle with basic skills.”
AI-driven learning technologies can support global equity in learning if we design them well, she said. The Allo Alphabet project, for example, is a phone-based literacy intervention in Côte d'Ivoire, a low-infrastructure setting where most households have mobile-phone access. A mobile app uses interactive voice-response to deliver literacy lessons and games to students at home and in school, with support from parents and teachers. Allo Alphabet has driven a large improvement in phonemic awareness and letter reading across users.
ClassInSight is a separate project, designed for higher-infrastructure education settings. Sensors collect within-classroom data on student and teacher movement and facial features and post information for all students from the teacher’s point of view. Teachers can use that information to understand who hasn’t spoken or had much attention, for example, and equalize education opportunities.
“It’s critical to ensure context-sensitive AI design incorporating both physical and human infrastructure,” Ogan said.
Accelerating Workplace Learning at Scale
Candace Thille, Amazon’s director of learning science, described AI as a Trojan horse for bringing science into the workplace-learning context.
“AI can support equity in any educational effort through intentional design, recognizing diversity among human learners related to place, skills, goals, background knowledge, and other factors,” Thille said.
Supporting learning-related diversity is especially critical at Amazon, which has 1.3 million employees and hires about 350 people every day. Moreover, the firm’s fast pace of innovation requires constant reskilling.
In this context, technology, including AI applications, means access, simulation capabilities, and connections to people — the three areas of workplace learning on which Thille focuses. “The workplace offers constant opportunity to understand the learner, what’s learned, and the context, to make instructional decisions.” AI-based models, specifically, can be used to observe learners and provide augmentative feedback.
In a recent experiment, Amazon used a reinforcement learning agent to set content-sequencing for employees learning linear algebra. Compared with a human instructor or self-navigation on the learners’ part, the agent performed best at improving learning outcomes most efficiently.
“But humans are still important parts of the process and the goal is not replacement,” Thille said. “The key is to look at the whole system: What parts can machines and algorithms support for decisions that have to be made, while protecting employees’ data and privacy?”