Research.

What This Lab Aims For

“We want what we learn to work on the next problem too.”

Have you ever applied logical thinking acquired in mathematics to programming, or used structured thinking developed in programming for essay writing? Insights gained from solving one problem carry over like interest to the next new problem. We call this “compound interest in problem-solving.”

Our research goal is to elucidate the conditions under which this compound interest arises and to realize them as working systems.

To do this, we use the concept of “intermediate representations.” Human ways of thinking and learning share common patterns that cut across domains, but as long as those patterns stay inside someone’s head, a computer cannot reason about them—so we describe the patterns in a form that computers can process. Once they are written down in the same framework, “correct knowledge” and “learner understanding” become directly comparable, which is what lets the system pinpoint where and how gaps exist instead of only judging a final answer right or wrong.

Four Research Projects

CHUNK: Making What You Learn Usable in Other Situations

Even after learning “for loops” in programming, you can’t use them in new problems. Even after memorizing mathematical formulas, you can’t solve application problems. “I learned it, but I can’t use it”—this is a barrier many learners experience, and it is a problem of knowledge transfer. How do we accumulate compound interest?

To overcome this barrier, we organize knowledge into three layers—"what it’s used for (function)," “how it works (behavior),” and “how to write it (structure).” We separate these layers because “I can’t use what I learned” usually means a learner memorized the structure (how to write it) without grasping the function and behavior, so when the surface form of a new problem differs, the knowledge no longer applies. When a learner understands “why it works” rather than just the wording, the same knowledge can be recognized and reused even when the new problem looks different—which is exactly how compound interest accumulates. A representative effort is Compogram, a learning environment where learners gradually build up component knowledge while visualizing the “behavior” of programs. Behavior is normally invisible inside a running program, and making it visible is what lets learners connect “how to write it” to “why it works.”

CLOVER: Learning Effectively from Mistakes

Conventional systems only judge “correct/incorrect,” but mistakes contain valuable information about “where understanding was ambiguous.” If there are gaps in the compound interest you’ve accumulated, they need to be corrected.

In the CLOVER project, we visualize wrong answers in the form of “what would happen if this were correct,” enabling learners to realize on their own, “Wait, this is different from what I expected.” We deliberately design experiences where learners notice for themselves rather than directly telling them “this is wrong,” because being told the correct answer fixes that one item but leaves the underlying misunderstanding untouched; when a learner sees for themselves that their idea leads to an impossible result, they revise the reasoning that produced the mistake, so the correction transfers to other problems too. By using intermediate representations, we can pinpoint not just “correct or incorrect” but “which part of their understanding has a gap.” A generic “try again” cannot target the specific gap, whereas an awareness experience aimed at the exact gap is far more likely to dislodge it. A representative effort is EBS (Error-based Simulation), an environment that promotes awareness of errors by showing simulations based on learners’ answers in subjects like mechanics.

OCEAN: Designing Optimal Learning for Each Individual

“I don’t know where to start studying.” “I’m not motivated today."—In today’s information-rich world, learners easily get lost. To maximize the effect of compound interest, each person needs a learning path suited to them.

In the OCEAN project, we develop adaptive learning environments that comprehensively grasp learners’ “understanding state,” “learning style,” “motivation,” and “goals,” and propose learning approaches suited to each individual. We look at all four together rather than just understanding because the reasons a learner stalls are not only cognitive—someone may understand the material yet still not study because they have lost motivation or cannot decide what to do next, and a recommendation that ignores those factors will be ignored in turn. Representative efforts include CORAL, a motivation support framework focusing on the emotion of “it’s too much trouble”—we single this feeling out since “it’s a hassle” is one of the most common reasons learning that should happen simply does not start; the ARK model that organizes “what to learn,” “what materials to use,” and “what to do,” the three questions a lost learner is usually stuck on, so that naming them separately lets the system see which one; and WHALE, an educational agent that uses ARK to make comprehensive recommendations.

CCS: Making “Design Knowledge” for Learning Support Shareable

“Programming education” and “writing instruction” actually deal with similar thinking skills, yet their systems are built separately. Using intermediate representations as a common language should reveal shared patterns, yet knowledge isn’t shared, and wheels are reinvented. Formalizing the mechanisms of compound interest themselves and making them shareable across domains—that is the aim of the CCS project.

We develop frameworks for describing “thinking skills” handled by different learning support systems in a common vocabulary, aiming to make support methods that were effective in one system reusable in others. Without such a vocabulary, the same idea is buried under different terms in each project, so even a method that clearly works in programming education stays invisible to writing instruction. Describe both in shared terms and the overlap becomes visible, and a proven method can be carried across instead of rediscovered from scratch in every field.

Research Approach

Our research proceeds by (1) understanding human thinking and learning, (2) theorizing it through intermediate representations as a medium, and (3) constructing and verifying theories as “working systems.” The key point is that system development is not mere “app development.” The system working correctly is itself a verification of the theory’s correctness. A system is built directly from the theory, so if the theory of how learning works is wrong, the system built on it will not produce the intended learning—and a system that actually works is evidence the theory behind it was right. This is also why we build at all rather than stopping at a written model: a theory left on paper can sound convincing while quietly being incomplete, and only making it run forces the gaps into the open.

Under this “understanding by building” approach, we conduct research across multiple fields including artificial intelligence, knowledge engineering, learning analytics, cognitive science, and educational psychology. Understanding human learning, formalizing it as knowledge a computer can use, and confirming it with real learner data each demand a different discipline; no single one of them can close the loop from “understanding people” to “a working system,” which is why we work across all of them rather than settle into one.

For more on the lab’s methodology and values, see About the Lab.

Acknowledgements

These research projects emerged from discussions with many colleagues, students, and collaborators. I am particularly deeply grateful to the following mentors who guided me: Takahito Tomoto, Tsukasa Hirashima, Tomoya Horiguchi, Hiroaki Ogata, Izumi Horikoshi, Rwitajit Majumdar, H. Ulrich Hoppe, Riichiro Mizoguchi, and Takako Akakura.