Five questions for Michael Mozer
Developing ways to optimize the power of learning, both for man and machine, has been a focus of Michael Mozer’s research for many years. Mozer, a professor in the Department of Computer Science and Institute of Cognitive Science (ICS) at CU Boulder, is studying the most effective ways of teaching so that humans can better consume and retain information. He’s also researching and testing tools that can make computers smarter and easier to use.
Mozer came to CU in 1988, when the university was one of the few schools in the country where faculty were doing research in neural networks, which are computer systems modeled on the human brain.
“Sadly, faculty in that area moved on, and for many years it was difficult to convince my colleagues that neural networks were a worthy area of research,” he said. “Now, of course, the AI (artificial intelligence) revolution is based on exactly the same neural network approaches we were exploring around 1990. You may have heard the term ‘deep learning,’ which is just a modern rebranding of neural nets. With faster computers and bigger data sets, the methods seem to work.”
1. How do the tools you are developing to improve human capabilities work? Why do humans – with our big brains – need them?
The best way to explain how the tools operate is by analogy to what online shopping sites do to make product recommendations. They collect data from a large population of customers purchasing a variety of products, and they use these data to recommend specific products to a particular individual. We do the same thing, except the recommendations concern what material a student ought to study, or incentives or guidance we might offer an individual to support better decision making. Our work involves three components: first, a means of quantifying optimality; second, large data sets of human behavior; and third, theories of human perception and cognition.
People need help learning because we have lapses of attention. A radiologist may miss detecting an anomaly on an X-ray. People have lapses of willpower: Students often skip class even when they know it is in their long-term interest to attend. And sometimes, we just have poor intuitions about how our minds work; for example, how and when to study material for long-term knowledge retention.
By the way, as an illustration of our poor intuitions, there is no solid evidence supporting the notion of different learning styles (Pashler, McDaniel, Rohrer, and Bjork, 2009), at least in the sense that people who consider themselves visual or auditory learners actually learn better with that style of instruction. It’s possible that if the style of instruction was chosen not based on an individual’s intuitions, but on the specific content of the material and specific history of the individual, there may actually be benefits. This is the type of question one can answer only with data and modeling.
2. Another project you are involved with is “temptation avoidance,” where incentives are used to keep people on track to meet long-term goals. How is this research done and what have you found?
Shruthi Sukumar did this work for her master’s thesis. We constructed a simple video game that simulated waiting in queues. If the player waits in a long, frustrating queue, they get a big payoff (per unit time), but if they jump to a short, painless queue, they get a smaller payoff. Even though the optimal strategy is to wait in the long queue, players often give up. After playing for a few minutes, we automatically construct a model of the player’s behavior, and using the model, we can predict an incentive structure that will maximize their expected payoffs. These incentives involve small rewards, kind of like offering a dieter a healthy snack at the moment they’re about to succumb to temptation. Although generic incentivizing schemes are used in behavioral economics (e.g., prize-linked savings accounts), our focus is on customizing incentives for an individual.
3. As you study ways to optimize human learning, one of your project focuses on helping achieve “durable memory.” What is durable memory and what does this research entail?
All human memory decays. Durable memories decay more slowly. The goal of our research is to slow the inevitable forgetting.
Pretty much everyone has heard the advice not to cram for an exam and that it’s better to distribute study over larger windows of time. That advice turns out to be wrong if your goal is to do well on the exam, but it’s right if you want to slow forgetting. We’re starting a project with researchers at Quizlet, an online flashcard app that has millions upon millions of users, predicting an individual’s rate of forgetting given their study history. Once we can model forgetting, we can use the model to prescribe optimal schedules of study in a precise, quantitative manner. We’ve done similar work on a smaller scale, with a group of 300 middle-school, Spanish-language students. A month after the semester ended, we found that using our adaptive training scheme achieves a nearly 17 percent boost in retention over a time-matched scheme that is representative of current educational practice.
4. One of your early projects was an “adaptive house” and you are currently working to gather data from digital textbooks. What is the status of each of those projects?
My 15 minutes of fame was a project in the 1990s where we equipped my house with sensors and actuators, and the house learned my habits and learned to control energy resources to both maximize my comfort and minimize energy consumption. We had a paper called the Adaptive Thermostat, written about 20 years before the Nest thermostat came along, and our system actually worked a lot better and was more intelligent than the Nest. The system also controlled lighting in a predictive manner, so that when you got up from the living room couch, it would anticipate — based on the context and an individual’s past behavior — that you were headed to the bedroom, and a lighting path would turn on. I kept the system running until there was a nasty lightning storm that fried most of the electronics. (To learn more, visit http://www.cs.colorado.edu/~mozer/index.php?dir=/Research/Projects/Adaptive%20house/)
The digital textbook project is just getting underway. We’re collaborating with an open-access textbook foundation (visit openstax.org) to instrument their electronic texts to collect data from students as they make initial contact with material. We record, for example, how long they look at a page and which sentences they highlight. The goal is to use these behaviors as early indicators of difficulties in understanding and particular topics of interest to that student. The textbooks will incorporate automatic quizzing, the purpose of which is to improve comprehension of the material.
My colleague Sidney D’Mello in ICS does related work. He has shown that when students’ minds wander, their reading times decouple from the intrinsic difficulty of the material. We hope to leverage relationships like this to optimize the student’s time and focus. (To learn more about the digital textbook project, visit http://tdlc.ucsd.edu/research/DNS/videos/Mozer.mp4 )
5. What aspect of your work do you enjoy most – figuring out how humans learn (think, etc.) or figuring out how to make machines (like your adaptive house) smarter? Why?
The two go hand in hand, in my view. It’s easy to make a machine smart at a specific task, such as detecting cats in YouTube videos, but machines do not have the flexibility and general cognitive abilities of people. We have a few examples where insights into the human mind and brain have proven useful for building smarter machines, but I think many more are to come.